Resources
The tech glossary: the vocabulary explained by recruiters.
AI, data, dev, cloud, defense, industry: 191 terms explained simply, with what to check when hiring.
191 terms
A/B Testing
Methodology= two versions, the numbers decide
A/B testing is an experimentation method that consists of showing two versions of the same element (page, feature, email, price) to two randomly assigned groups of users, then comparing which one best achieves a measured goal, such as conversion rate or retention.
The principle is simple, the rigor less so. You need to define the metric before launching the test, calculate the required sample size and wait for enough data for the observed difference to be statistically significant, not due to chance. Stopping a test as soon as it "looks like it's winning" is the most common mistake.
Technically, tests often rely on feature flags to show one variant or the other. Tools such as Optimizely, GrowthBook or Statsig handle assignment and analysis. It is a core tool for product and growth teams.
Agile
Methodology= moving forward in small, checked steps
Agile is a project management approach, born in software, that consists of delivering in short cycles, gathering regular feedback and adjusting the plan continuously rather than locking everything in at the start. It stands in contrast to the "waterfall" model, where specification, development and testing happen in large successive phases.
The founding text is the Agile Manifesto, signed in 2001 by 17 software development practitioners. It sets out four values, including priority given to individuals and interactions, working software and responding to change.
In practice, Agile comes in frameworks: Scrum (one to four week sprints, defined roles), Kanban (continuous flow, limits on work in progress), Extreme Programming (engineering practices such as TDD), or scaled frameworks like SAFe. The word is now used far beyond tech, sometimes in a very diluted way.
AI
TechConcept= the machine that mimics reasoning
Artificial intelligence (AI) refers to the set of techniques that let a machine perform tasks associated with human intelligence: recognizing an image, understanding a text, making a decision, generating content. The term was proposed by John McCarthy for the 1956 Dartmouth conference.
For decades, AI relied mainly on hand-written rules (expert systems). Since the 2010s, it has relied mostly on machine learning: the model learns from examples rather than rules. Deep learning, then Transformer architectures, paved the way for large language models (LLMs) and generative AI.
Today a distinction is drawn between "predictive" AI (scoring, anomaly detection, vision) and "generative" AI (text, code, images). Both coexist in industry, healthcare, defense and finance, as well as in everyday software.
AI Act
TermConcept= the highway code for AI
The AI Act is the European Union regulation on artificial intelligence, adopted in 2024, which governs placing AI systems on the market and using them according to the level of risk they present. It applies to any organization that provides or uses an AI system in the EU, regardless of its country of origin.
It distinguishes several levels:
- prohibited practices, such as social scoring or certain forms of manipulation;
- high-risk systems (recruitment, education, credit, critical infrastructure, etc.), subject to documentation, human oversight and data quality obligations;
- limited risk, with transparency obligations (disclosing that users are talking to an AI, labeling generated content);
- minimal risk, with no specific obligation.General-purpose AI models, such as LLMs, have their own obligations. Application is phased in over several years, and fines can be calculated as a percentage of worldwide revenue.
AI Agent
TechConcept= AI that acts, not just answers
An AI agent is a program that relies on a language model (LLM) to reach a goal in several steps: it decides which actions to take, uses tools (search, code execution, API calls, reading and writing files), observes the result and adjusts what comes next, with more or less human supervision.
The difference from a chatbot: a chatbot answers a question, an agent chains actions to complete a task. A coding agent, for example, can read a repository, edit files, run the tests and fix its own mistakes. The open MCP protocol (Model Context Protocol) is used to connect tools and data sources to these agents in a standard way.
Building a reliable agent is hard: you have to limit its permissions, add guardrails, measure the quality of its output and manage the cost and latency of many model calls.
AI Engineer
Role= the developer who plugs in AI
The AI Engineer is the engineer who builds products and features on top of existing AI models, especially large language models (LLMs), and puts them into production reliably. They generally do not train models: they assemble them, connect them to company data and measure their quality.
Their work covers RAG (connecting a model to a document base), prompt engineering, designing agents and tool calls, automated evaluation of answers, observability, and managing cost and latency. They mostly use Python or TypeScript, model provider APIs (Anthropic, OpenAI, Mistral…) or open weights models, and vector databases.
The profile often comes from back-end or full-stack development rather than research.
Airflow
ToolTech= the conductor of data pipelines
Apache Airflow is an open source orchestration platform used to schedule, run and monitor chains of data tasks, called pipelines. It was created at Airbnb in 2014, then handed over to the Apache Software Foundation.
Pipelines are written in Python as DAGs (Directed Acyclic Graphs): a graph of tasks and their dependencies (extracting data, transforming it, training a model, producing a report). Airflow runs each task at the right time, in the right order, retries those that fail and provides an interface to monitor everything.
It is a reference tool in data engineering, often paired with dbt for transformations and a warehouse such as Snowflake or BigQuery. It has competitors (Dagster, Prefect) and is available as a managed service from the major cloud providers.
Angular
FrameworkTech= the all-in-one front-end toolkit
Angular is an open source front-end framework, maintained by Google, used to build rich web applications, especially single-page applications (SPAs) where the interface updates without reloading the page. It is written in TypeScript.
Angular is a complete rewrite of AngularJS, its JavaScript predecessor, which is no longer maintained. The two are incompatible: migrating from one to the other almost amounts to rewriting the application.
Its distinctive feature is being "batteries included": components, routing, forms, HTTP calls, dependency injection and testing tools are provided and structured by the framework. Where React leaves many choices open, Angular imposes an organization. That explains its strong presence in large corporations, banking, insurance and internal business applications, where code consistency across big teams matters more than freedom.
API
TermConcept= the power socket between two programs
An API (Application Programming Interface) is a set of rules that lets one piece of software request data or actions from another, without knowing how it works internally. It defines what can be requested, in what form, and what comes back.
Example: when a weather app shows the temperature, it queries a data provider's API. When a website accepts a payment, it calls the API of a provider such as Stripe. The application does not need to know how those services are built.
On the web, most APIs run over HTTP. The most common styles are REST (resources reachable through URLs), GraphQL (the client asks for exactly the fields it needs) and gRPC (fast, mostly between internal services). An API can be internal to a company, open to partners or public, free or paid. It then becomes a product in its own right, with its documentation, versioning and quotas.
Astro
FrameworkTech= fast sites without needless JavaScript
Astro is an open source web framework designed for content-heavy sites (marketing sites, blogs, documentation, e-commerce), which by default generates HTML pages without sending JavaScript to the browser, for pages that load very fast.
Its key principle is the islands architecture: the page is static, and only the components that need interactivity (a carousel, a form, a cart) load their JavaScript, independently of one another. These islands can be written with React, Vue, Svelte or other libraries, in the same project.
Astro can produce a fully static site or render pages on the server. It is often chosen for marketing sites where performance and search ranking matter more than the interactivity of a full application.
ATS
Tool= the recruiting dashboard
An ATS (Applicant Tracking System) is the tool that centralizes a hiring process: posting jobs, receiving and screening applications, tracking each candidate stage by stage, communicating with managers and reporting.
In practice, each open role becomes a pipeline in columns (application, HR interview, technical test, final interview, offer). The ATS keeps the history of exchanges, schedules interviews, sends replies and extracts information from resumes (parsing). Many now offer AI-assisted matching or writing features.
There are ATSs for companies, designed for in-house HR teams and often paired with a careers page (Greenhouse, Lever, Teamtailor, Workday), and tools for recruitment agencies, which add a CRM layer to manage clients and assignments. The choice depends on hiring volume and the number of people involved.
Back-end
Term= the application's engine room
The back end is the part of an application that runs on servers, invisible to the user: it receives requests from the interface, applies business rules, reads from and writes to databases and sends back results. It is the counterpart of the front end, the part displayed in the browser or mobile app.
When you check out an online cart, the back end checks stock, computes the price, calls the payment service, records the order and triggers the confirmation email.
Back-end development covers API design, data modeling, security, performance and integration with other systems. Common languages are Java, Python, Go, PHP, Ruby, C# or JavaScript and TypeScript through Node.js. The associated frameworks (Spring, Django, Symfony, Laravel, Rails) give this work its structure.
Blockchain
TechConcept= the tamper-proof shared ledger
A blockchain is a ledger of transactions shared across many computers, where data is grouped into blocks chained together through cryptography, which makes any later modification extremely difficult. No central authority keeps the ledger: the network participants validate additions.
Validation relies on a consensus mechanism. Proof of work (Bitcoin) requires solving energy-intensive computations; proof of stake (Ethereum since 2022) relies on participants who lock up tokens as collateral.
The first use case is Bitcoin, launched in 2009. Ethereum added smart contracts, programs executed on the blockchain, which enabled DeFi, NFTs and DAOs. There are also private blockchains, between identified partners, for traceability or interbank exchanges. The main challenges remain cost, speed (hence "layer 2" solutions) and regulation.
Bun
TechTool= the fast all-in-one Node.js
Bun is a JavaScript and TypeScript runtime that packs a runtime, an npm-compatible package manager, a bundler and a test runner into a single binary.
It is written in Zig and built on JavaScriptCore, Safari's JavaScript engine, whereas Node.js and Deno use V8. Created by Jarred Sumner, it reached version 1.0 in September 2023.
Its main selling point is speed: fast startup, very short dependency installs, and direct execution of TypeScript files with no compile step. It aims for compatibility with Node.js APIs, so a large share of existing code runs without being rewritten.
You mostly find it in development tooling (scripts, tests, package installs) and in newer back-end services, less often as the main production runtime of an older application.
Chaos Engineering
MethodologyConcept= breaking things on purpose
Chaos engineering is a discipline that deliberately injects failures into a system (a server shut down, a slowed network, an unavailable dependency) to check that it holds up as expected, before a real incident exposes the weakness.
The approach is experimental: you state a hypothesis (“if a cloud zone goes down, the service stays available”), run the experiment on a limited scope, measure it with observability, then fix whatever gave way. Experiments start in pre-production, then move to production once the team is confident.
The practice was popularized by Netflix and its Chaos Monkey tool, which randomly shut down production instances. Tools such as Gremlin, Chaos Mesh, LitmusChaos and AWS Fault Injection Service have made it more accessible.
It is mostly found in SRE teams running high-availability services: payments, consumer platforms, critical infrastructure.
CI / CD
MethodologyConcept= the automated assembly line for code
CI/CD (Continuous Integration / Continuous Delivery or Deployment) is the automation of the path that takes a code change to production: each change is automatically built, tested, then prepared or deployed without manual handling.
Continuous integration (CI) means merging developers' work into the shared codebase very often; at each merge, a suite of automated tests checks that nothing is broken. Continuous delivery (CD) ensures that validated code is always ready to go to production in one click. Continuous deployment goes further: it ships to production automatically if all tests pass.
These steps are chained in a pipeline, configured in tools such as GitHub Actions, GitLab CI, Jenkins or CircleCI. The benefit: frequent, small and less risky releases, instead of large, stressful quarterly deliveries.
Claude
Tech= Anthropic's AI assistant
Claude is the family of large language models (LLMs) developed by Anthropic, a US company founded in 2021 by former OpenAI staff and focused on AI safety.
The models come in several tiers, from the fastest and cheapest to the most capable (Haiku, Sonnet, Opus). They are known for writing code, analyzing long documents and using tools inside agents. Their training relies in particular on Constitutional AI, a method in which the model learns to follow a list of explicit principles.
Claude is available through the claude.ai app, Anthropic's API and the major clouds (AWS Bedrock, Google Cloud Vertex AI). Anthropic also offers Claude Code, a tool for handing development tasks to Claude directly in the terminal or the editor.
Anthropic also created MCP (Model Context Protocol), an open standard for connecting models to tools and data.
Cloud Act
TermConcept= the US law that follows data
The Cloud Act (Clarifying Lawful Overseas Use of Data Act) is a US law passed in 2018 that allows US authorities to require a provider subject to US law to give access to data it controls, including when that data is stored outside the United States.
The criterion is therefore not where the server is, but the nationality or legal ties of the company operating the service. Hosting data in a datacenter located in France with a US provider is not enough to put it beyond the reach of this law. The text provides for ways to challenge requests and for bilateral agreements between states.
The Cloud Act is at the heart of the European debate on digital sovereignty. It is often compared with the GDPR and led to the addition, in the SecNumCloud framework, of criteria for immunity from non-European law.
Cloud computing
TermTech= IT rented on demand
Cloud computing is the provision of IT resources (servers, storage, databases, software) on demand over the Internet, billed by usage rather than bought and installed on site.
Instead of buying and hosting machines, a company rents capacity from a provider and adjusts it to its needs. There are three main levels: IaaS (raw infrastructure), PaaS (a platform to deploy your code) and SaaS (ready-to-use software). The cloud can be public, private or hybrid.
The cloud made it possible to build digital products quickly without upfront hardware investment. In return, it raises questions about long-term cost, dependence on a provider and data location, hence the rise of FinOps and sovereign cloud offerings.
Compute
Term= the available computing power
Compute refers to the computing power used to run software, and in AI more specifically the resources (GPUs, accelerators, servers) needed to train and serve models.
The word moved from engineering jargon to executive vocabulary with the rise of large language models. A model's capabilities improve largely with the amount of compute invested in training, which makes it both a major expense and a strategic issue. People talk about "access to compute" the way they would talk about access to capital.
In practice, compute is bought (in-house servers and datacenters) or rented (cloud, specialized GPU providers). AI companies constantly trade off training compute, to build or adapt a model, against inference compute, to serve it to users.
Container
TermTech= the application packed in its suitcase
A container is a software package that holds an application with everything it needs to run (libraries, dependencies, configuration), and that runs in isolation on any machine equipped with a container engine. It solves the classic "it works on my machine, not in production".
Unlike a virtual machine, which ships a complete operating system, a container shares the host system's kernel. It is therefore much lighter and starts in seconds, which makes it possible to run dozens of them on the same server.
Docker popularized the format from 2013 onwards; the standard is now open (OCI). When managing hundreds of containers, teams use an orchestrator such as Kubernetes, which spreads them across servers, restarts them on failure and adjusts their number to the load. Containers have become the standard way to deploy applications in the cloud.
Context Window
TermConcept= the model's working memory
The context window is the maximum amount of text a language model can process at once, measured in tokens: instructions, conversation history, supplied documents and the generated answer all included.
Anything beyond that limit is ignored, or has to be summarized, split up or retrieved on demand. Windows have grown a lot: some models go beyond a million tokens, enough to hold several books or an entire codebase.
A large window does not solve everything. The more you load into the context, the more each call costs and the longer it takes, and models often make poorer use of information placed in the middle of a long text (the lost in the middle effect). Hence the value of RAG and careful context management: give the model only what is useful.
The topic is central for AI agents, which quickly pile up tool results and must sort out what they keep in memory.
Copilot
ToolTerm= AI suggests, the human approves
A copilot is an AI-based assistant built into a piece of software that suggests code, text or actions in real time, which the user accepts, edits or rejects.
The term comes from GitHub Copilot, launched in 2021 by GitHub (a Microsoft subsidiary) with OpenAI: it completes code directly in the editor. Microsoft then extended the Copilot brand to its other products, from Windows to the Microsoft 365 suite. The word has become a generic name for any assistant of this kind.
In software development the offer is broad: GitHub Copilot, Cursor, JetBrains AI Assistant, Claude Code, Windsurf, Tabnine… The line with agents is blurring: these tools no longer just complete a line, they can edit several files or run commands.
The distinction is still useful: a copilot suggests and the human approves each step, while an agent chains several actions on its own toward a goal.
CRM
Tool= the memory of customer relationships
A CRM (Customer Relationship Management) is the software where a company centralizes everything about its customers and prospects: contact details, history of exchanges, open sales opportunities, contracts, support tickets.
For a sales team, it is the daily working tool: each deal moves through a pipeline (first contact, meeting, proposal, signature), which makes it possible to forecast revenue and never miss a follow-up. For marketing, it is used to segment and measure campaigns. For support, it gives a customer's context at a glance.
The best known are Salesforce, HubSpot, Pipedrive and Microsoft Dynamics. In a growing company, the CRM quickly becomes a central part of the information system, connected to billing, emailing and dashboards. Its quality depends above all on how rigorously teams enter data.
CSS
Language= the styling of web pages
CSS (Cascading Style Sheets) is the language that defines how web pages look: colors, fonts, spacing, layout, animations. HTML describes the content and its structure; CSS decides how it is displayed.
The "cascade" principle means that several rules can apply to the same element, with the most specific one winning. It is powerful, but a source of conflicts on large projects if the code is not organized.
Modern CSS makes complex layouts possible with Flexbox and Grid, and adapts to every screen size through media queries (responsive design). To stay organized, teams use naming conventions, preprocessors such as Sass, utility frameworks such as Tailwind CSS or styles attached to components. CSS is maintained by the W3C and evolves continuously.
CTO
Role= the person in charge of technology
The CTO (Chief Technology Officer) is the executive responsible for a company's technology: technical choices, architecture, organizing and hiring the tech team, security and the ability to deliver the product. They sit on the executive committee.
The role changes enormously with the company's stage. In a startup of a few people, the CTO is often a co-founder and codes a good part of the product. As the team grows, the focus shifts to hiring, structuring the team and product trade-offs. In a large company, the CTO carries the technology vision and the relationship with the business, while team management goes to a VP of Engineering.
In industry or deeptech, the CTO also owns the scientific roadmap and intellectual property (patents, R&D). In a corporate group, the role should not be confused with the CIO, who focuses on the internal information system.
Cybersecurity Analyst
Role= the watchkeeper of security alerts
A Cybersecurity Analyst is the professional who monitors an organization's information systems, detects threats, analyzes security incidents and takes part in the response.
Day to day, they work in a SIEM (a tool that centralizes logs and alerts) and in detection tools on workstations and servers (EDR). They triage alerts, rule out false positives, investigate suspicious cases, assess an incident's severity and coordinate the first actions with IT teams.
They usually sit in a SOC (Security Operations Center), in-house or at a managed security services provider. Levels are often split into L1 (alert triage), L2 (investigation) and L3 (incident response, threat hunting).
The role exists in every sector, with growing demand among operators subject to security obligations such as NIS2.
DAO
TermTech= an organization run by code
A DAO (Decentralized Autonomous Organization) is a collective whose operating rules are written into smart contracts on a blockchain, and whose decisions are voted on by holders of governance tokens.
In practice, a member submits a proposal (funding a project, changing a protocol parameter), token holders vote, and the result can be executed automatically by the contract. A DAO often manages a shared treasury in crypto-assets.
DAOs are mostly used to govern DeFi protocols, fund open source projects or public goods, or pool investors. The 2016 episode of “The DAO”, partly drained of its funds through a flaw in its contract, is a reminder that code is law, bugs included.
The limits are well known: unclear legal status in many countries, often low voter turnout, and power concentrated among large token holders.
Data
Term= the trace left by every action
Data is information recorded in a form that a computer system can store, process and transmit: an order amount, a GPS position, a click, a sensor reading, a text, an image.
Every digital action produces some: a payment, a search, a message, a visit to a web page. In industry, machines and sensors generate it continuously. A distinction is made between structured data, organized in tables (a customer database), and unstructured data, such as emails, documents or videos, which are harder to use.
Raw data is worth little. Its value comes from the chain that collects, cleans, stores and makes it usable for decisions or automation: dashboards, predictive models, AI training. This chain involves several jobs (Data Engineer, Data Analyst, Data Scientist) and a strict legal framework as soon as personal data is involved (GDPR).
Data Analyst
Role= the translator of numbers into decisions
The Data Analyst is the person who turns a company's data into concrete answers to business questions: why sales are dropping in a region, which acquisition channel pays off most, where users drop out of the journey.
Their daily work: extracting and combining data in SQL, cleaning it, building dashboards (Looker, Power BI, Tableau, Metabase), then presenting conclusions to product, marketing, finance or leadership teams. They also use Python or an advanced spreadsheet for one-off analyses.
They sit between the Data Engineer, who builds and secures data flows upstream, and the Data Scientist, who designs predictive models. In small teams, the boundaries are blurry and the Analyst often handles part of the data modeling, a role sometimes called Analytics Engineer when it relies on dbt.
Data Contract
TermMethodology= the spec sheet for data
A data contract is an explicit, versioned agreement between the team that produces a dataset and the teams that consume it, setting out its schema, its meaning, its update frequency and its quality rules.
The contract describes the fields, their types, the allowed values, the expected delivery time and the procedure for breaking changes. It is written in a machine-readable format (often YAML) and checked automatically: a change that breaks the contract is blocked before it reaches consumers.
The problem it solves is common: an application team renames a column, and the dashboards or machine learning models that depended on it break silently.
It is found in mature data organizations, often alongside a data mesh approach. It relies on standards such as the Open Data Contract Standard, data quality testing tools (Soda, Great Expectations) or a schema registry for Kafka streams.
Data Engineer
Role= the plumber of data flows
The Data Engineer is the engineer who builds and maintains a company's data infrastructure: they collect data from applications, tools and sensors, transform it, store it and make it available, reliable and up to date, for analytics and AI.
In practice, they design pipelines that run daily or in real time, orchestrated by tools such as Airflow, transformed with dbt or Spark, and stored in a warehouse or lakehouse (Snowflake, BigQuery, Databricks). They monitor data quality, manage compute costs and access security.
Their work sits upstream of the Data Analyst's and the Data Scientist's: without solid pipelines, their analyses and models rest on wrong or incomplete data. They are close to back-end developers in their practices (code, tests, CI/CD) and to infrastructure engineers in their cloud tooling.
Data Lakehouse
TechTerm= a tidy data lake
A data lakehouse is a data architecture that combines the cheap, open storage of a data lake (files in S3-type object storage) with the guarantees of a data warehouse: reliable transactions, schemas and fast SQL queries.
The key is an open table format (Apache Iceberg, Delta Lake or Apache Hudi) layered on top of Parquet files. This metadata layer brings transactions, the ability to query a past version of the data (time travel) and schema evolution without rewriting everything.
The benefit: a single copy of the data serves BI analysis, data science and machine learning alike, and several engines (Spark, Trino, Snowflake, DuckDB…) can read the same tables. This limits dependence on a single vendor.
The term was popularized by Databricks. It is found in companies handling large volumes or wanting to keep control of their storage.
Data Mesh
MethodologyTerm= data handed to business teams
Data mesh is an approach to organizing data, formalized by Zhamak Dehghani in 2019, that gives responsibility for data to each business domain rather than to a single central data team.
It rests on four principles. Each domain owns and maintains its data. That data is treated as a product, with an owner, documentation and quality commitments. A shared self-service platform saves each team from rebuilding everything. Finally, federated governance sets the shared rules (security, formats, interoperability).
Data mesh addresses the bottleneck of the central data team, swamped with requests about data it knows poorly. In practice, it is tooled with data contracts, a data catalog and an internal data platform.
It is first and foremost an organizational change. It makes sense in large companies with many domains, much less in a company with a single data team.
Data Scientist
Role= the statistician who makes data talk
The Data Scientist is the specialist who uses statistics and machine learning to draw forecasts, explanations or automated decisions from data: forecasting demand, detecting fraud, estimating the risk that a customer will churn, recommending a product.
They frame the problem with business teams, explore the data, choose and train a model (Python, scikit-learn, PyTorch), then measure whether it beats a simple rule. A large part of the work consists of validating results and explaining their limits.
The job has specialized. Putting models into production and monitoring them often falls to the ML Engineer or MLOps Engineer, and integrating existing LLMs to the AI Engineer. The Data Scientist keeps the analytical and statistical core: experimentation, causality, custom modeling. In deeptech and industry, they also work on sensor, image or simulation data.
Database / DB
TechTerm= the application's organized memory
A database (DB) is a set of information stored in a structured way so that it can be found, modified and protected efficiently, managed by software called a DBMS (database management system). An online store's customers, their orders and the product catalog all live in a database.
It is queried with queries. Relational databases store data in linked tables and are queried in SQL: PostgreSQL, MySQL, SQLite, Oracle, SQL Server. They guarantee transaction consistency, which is essential in finance or logistics.
NoSQL databases drop the table model for other forms: documents (MongoDB), key-value (Redis), wide columns (Cassandra), graphs (Neo4j). They serve specific needs such as an ultra-fast cache or highly distributed volumes. On top of these come analytical data warehouses (Snowflake, BigQuery) and the vector databases used by AI.
dbt
ToolFramework= SQL organized like code
dbt (data build tool) is an open source tool for transforming data directly inside a data warehouse by writing SQL, with software engineering practices: Git versioning, tests, documentation and automated deployment.
A dbt project is a set of models, SQL queries that each build a table or view from others. dbt works out the execution order from the dependencies between models, compiles everything (with Jinja macros) and runs it in the warehouse: Snowflake, BigQuery, Databricks, Redshift, PostgreSQL… Tests check the data (uniqueness, null values, relationships) and documentation is generated from the project.
It is developed by dbt Labs, with an open source version (dbt Core) and a cloud offering.
dbt sits at the heart of what is called the modern data stack and gave rise to the analytics engineer role, between the data engineer and the data analyst.
DDD
Methodology= code that speaks the business language
DDD (Domain-Driven Design) is a software design approach that puts the business at the center: the code faithfully reproduces the concepts, rules and vocabulary of the domain (insurance, logistics, payments), built in close dialogue with domain experts. The approach was formalized by Eric Evans in a book published in 2003.
Two ideas are at its core. The ubiquitous language: developers and experts use the same words, and those words appear in the code. Bounded contexts: a large system is split into coherent zones, each with its own model; a "customer" does not mean the same thing for billing and for support.
DDD also offers tactical building blocks (entities, aggregates, value objects, domain events). It is useful for software with high business complexity and often serves as a guide for splitting an application into microservices.
Deep Learning
TechConcept= learning through successive layers
Deep learning is a branch of machine learning that uses artificial neural networks with many layers to learn directly from raw data: images, sounds, texts, signals. Each layer extracts increasingly abstract features, from the edges in an image up to the whole object.
These networks are very loosely inspired by the brain, but remain mathematical functions whose parameters are adjusted during training, using large amounts of examples and substantial computing power, usually GPUs.
The boom dates back to the early 2010s, when these methods clearly outperformed classic approaches in image recognition. Deep learning underpins computer vision, speech recognition, machine translation and, with the Transformer architecture, large language models. The reference frameworks are PyTorch and, to a lesser extent, TensorFlow and JAX.
Defense security clearance
TermConcept= permission to access secrets
A defense security clearance is the authorization granted by the state to a person to access classified information or materials, at a given level and for the needs of their mission. In France, since the 2021 reform, there are two levels: Secret and Très Secret (Top Secret).
The application is not made by the candidate themselves: it is filed by the employer, or the entity that needs the person to access classified information, and it is tied to a position. It triggers an administrative investigation carried out by the competent services, the DRSD for defense companies. The "need to know" principle applies: being cleared does not grant access to everything.
Clearances are found at defense manufacturers, in space, intelligence, some cybersecurity activities and among their subcontractors. A clearance can be denied, and it ends or must be reviewed when the person changes roles.
DeFi
TermTech= finance without banks, in code
DeFi (Decentralized Finance) refers to financial services (trading, lending, borrowing, derivatives) that run on smart contracts on a public blockchain, with no bank or central intermediary.
The basic building blocks are decentralized exchanges (DEXs) such as Uniswap or Curve, lending protocols such as Aave or Compound, and stablecoins, tokens whose value is pegged to a currency such as the dollar. Ethereum and its Layer 2s account for a large share of the activity.
Each protocol is a set of public contracts that anyone can use or combine with others. This openness is also its weakness: a flaw in a contract can cause immediate, irreversible losses, and protocol hacks have been frequent.
In Europe, the MiCA regulation covers crypto-assets and their service providers, but fully decentralized DeFi remains largely outside its scope.
Deno
TechTool= Node.js rethought by its creator
Deno is an open source JavaScript and TypeScript runtime created by Ryan Dahl, the author of Node.js, and introduced in 2018 as a fix for the design choices he regretted in Node.
It is written in Rust and built on the V8 engine. Three choices set it apart. Security by default: a script can only access the network, files or environment variables if explicitly allowed. TypeScript works with no configuration. And it follows browser standards (fetch, ES modules, Web APIs).
Deno also ships its own built-in tools: formatter, linter, test runner. Version 2, released in 2024, significantly improved compatibility with Node.js and npm packages, which makes gradual adoption easier.
The company behind Deno also offers Deno Deploy, a platform for running code at the edge. It is found in scripts, internal tools and lightweight back-end services.
Design pattern
Term= the proven recipe for a common problem
A design pattern is a standard, proven and reusable solution to a software design problem that comes up often. It is not ready-to-copy code, but an organizational scheme that you adapt to your context.
The notion was popularized in 1994 by the book Design Patterns by four authors nicknamed the "Gang of Four", which describes 23 patterns grouped into three families: creational (how to build objects, for example Singleton or Factory), structural (how to assemble components, for example Adapter or Decorator) and behavioral (how they communicate, for example Observer or Strategy).
Their first benefit is a shared vocabulary: saying "we go through an Observer" is enough for a team to understand the intent. Used well, they make code more readable and easier to evolve. Used badly, they add pointless layers of abstraction. There are also architecture patterns (MVC, CQRS) at a broader scale.
Developer Advocate / DevRel
Role= the developers' spokesperson
A Developer Advocate (or DevRel, for Developer Relations) is a technical profile responsible for getting developers to adopt a product and for bringing their needs back to the product and engineering teams.
The role works both ways. Outward: writing tutorials and code samples, improving documentation, speaking at conferences or meetups, running a community on GitHub, Discord or forums. Inward: relaying user feedback, flagging friction points, testing new features the way a customer would.
The role mostly exists at vendors whose customers are developers: APIs, SDKs, databases, cloud tools, AI models. For them, adoption depends largely on the quality of the developer experience.
Depending on the company, DevRel reports to product, marketing or engineering, which changes its goals significantly.
DevOps
RoleMethodology= the bridge between those who code and those who operate
DevOps is both a culture and a job. The culture aims to bring development (Dev) and operations (Ops) teams closer together in order to ship more often and more safely. The DevOps engineer's job is to build and automate the chain that takes code to production, and the infrastructure that runs it.
The movement emerged around 2009, notably with the first DevOpsDays conferences, in reaction to the walls between developers who wanted to ship fast and operations teams in charge of stability.
Day to day, the DevOps engineer sets up CI/CD pipelines, describes infrastructure as code (Terraform), manages containers and orchestrators (Docker, Kubernetes), monitoring and the security of cloud environments. The title overlaps with other roles: the SRE focuses on the measured reliability of services, the Platform Engineer on a self-service internal platform for developers.
Digital sovereignty
Concept= keeping control of your technology
Digital sovereignty is the ability of a state, a company or an organization to control its technologies, data and digital infrastructure, without depending on foreign providers or laws it does not control.
It plays out at several levels: where data is hosted and under which law (cloud, Cloud Act), who supplies critical components (chips, software, AI models), and who holds the skills to evolve them. It does not mean producing everything yourself, but being able to choose and switch providers without being locked in.
The topic spans defense, public administration, healthcare, industry and AI. It shows up in SecNumCloud-qualified cloud offerings, in public procurement policies and in the growth of European open-weight AI models.
Django
FrameworkTech= the complete web kit in Python
Django is an open source web framework written in Python, designed for quickly developing robust web applications. It was born at a local Kansas newspaper, the Lawrence Journal-World, and released as open source in 2005. It is maintained by the Django Software Foundation.
Its philosophy is "batteries included": out of the box it provides an ORM to work with the database without writing SQL, a URL system, templates, forms, user management and default security protections. Its automatically generated admin interface is one of its best-known strengths.
Django follows a split close to MVC, which it calls MTV (model, template, view). It is often paired with Django REST Framework to build APIs. It is found in SaaS products, content platforms, internal tools and many data projects, thanks to its closeness to the Python ecosystem.
DO-178C
TermMethodology= the standard for software that flies
DO-178C is the reference standard for certifying software embedded in aircraft. It defines the objectives a software development must meet to be accepted by certification authorities, with five levels of rigor, from DAL A to DAL E.
The level (Design Assurance Level) depends on the consequences of a failure: DAL A for software whose failure can be catastrophic, such as flight controls, down to DAL E when it has no effect on safety. The higher the level, the stronger the requirements for traceability, reviews, testing and code coverage.
Published by RTCA in the United States and EUROCAE in Europe (as ED-12C), it comes with supplements on formal methods, modeling and tools. It is found at aircraft manufacturers, equipment suppliers and their subcontractors, and it inspires other critical sectors.
Docker
TechTool= the tool that puts applications in boxes
Docker is an open source platform that packages an application and all its dependencies into a container, then runs it identically on a developer laptop, a server or in the cloud. It was introduced in 2013 by Solomon Hykes, at the company dotCloud.
It relies on three notions. The Dockerfile describes, step by step, how to build the application's environment. The image is the result of that build, a frozen template that can be shared through a registry such as Docker Hub. The container is a running image, isolated from the rest of the machine.
Docker brought containers to every developer: it simplifies setting up a project, guarantees consistency between environments and serves as the basis for CI/CD pipelines. To run many containers in production, teams rely on an orchestrator such as Kubernetes. Docker Compose, for its part, is used to start several containers together locally.
E-commerce
Term= the store open online
E-commerce (electronic commerce) refers to selling goods or services online: on a merchant website, a mobile app, a marketplace or social networks, with ordering and payment done over the internet.
Several models exist: B2C (a brand sells to consumers), B2B (between businesses, with catalogs and negotiated prices), D2C (a brand sells directly without a distributor) and marketplaces that connect third-party sellers with buyers.
Behind the storefront, an e-commerce site is a complete system: product catalog, search engine, cart, payment, inventory management, logistics, customer service, and analytics and marketing tools to attract and retain customers. Many merchants rely on off-the-shelf solutions (Shopify, Adobe Commerce, PrestaShop, WooCommerce); the largest often build their own platform. Site performance and payment reliability directly affect revenue.
Edge Computing
TechConcept= computing close to the action
Edge computing means running processing as close as possible to where data is produced or consumed, rather than in a centralized cloud datacenter.
The term covers two worlds. On the web, it refers to code running on points of presence spread around the globe (Cloudflare Workers, Vercel, Fastly, Deno Deploy) to respond faster to the user: routing, authentication, personalization. These environments are more constrained than a regular server, often with no file system access.
In industry, IoT, defense or vehicles, it refers to computing on the machine, the sensor or a local gateway: a robot, a camera or a drone that analyzes its data on the spot. The reasons are latency, limited or missing network connectivity, transfer costs and data confidentiality.
Both uses share the same principle: move computing to the data, and accept more limited resources in return.
Embedded system
TermTech= the computing hidden inside machines
An embedded system is an electronic and software assembly built into a piece of equipment to perform a specific function, often under strong constraints on size, power consumption, reliability and real-time behavior.
They are everywhere: car control units, aircraft flight controls, satellites, drones, medical devices, industrial robots, connected objects. The core is usually a microcontroller, a processor or an FPGA, programmed in C, C++ or increasingly Rust, sometimes with a real-time operating system. The software has to talk directly to the hardware: sensors, communication buses, actuators.
In critical sectors (aerospace, defense, rail, medical), development follows strict safety standards such as DO-178C or ISO 26262, with extensive testing and traceability.
Embeddings
TermConcept= meaning turned into coordinates
An embedding is a representation of a piece of content (text, image, sound) as a vector, a list of several hundred or thousand numbers, computed by a model so that content with similar meaning gets similar vectors.
In practice, “employment contract” and “permanent job” will have neighboring vectors even if they share no words. This closeness is measured by computing a distance between vectors, which makes it possible to search by meaning rather than by keywords.
Embeddings are used for semantic search, recommendation, document clustering, duplicate detection and above all RAG: they find the passages in a document base most relevant to a question before passing them to the LLM.
They are produced by specialized models, offered by AI providers or available as open source, and stored in vector databases (Qdrant, Weaviate, Pinecone, pgvector for PostgreSQL…).
Engineering Manager
Role= the manager of engineers
The Engineering Manager (EM) is the manager of a team of software engineers. They are responsible for the people (hiring, one-on-ones, skill development, evaluation) and for the team's ability to deliver reliably and sustainably.
They do not make technical choices alone: those belong rather to the tech lead or senior engineers, with whom they work in tandem. Nor do they decide what to build, which is the Product Manager's role. Their ground is organization: prioritization with product, delivery processes, dependency management, team climate and departures to anticipate.
The role appears when a tech team grows to the point where the CTO or lead developer can no longer follow each person. Above them sits a Head of Engineering or a VP of Engineering, who manages several EMs.
FDD
Methodology= development one feature at a time
FDD (Feature-Driven Development) is an iterative software development method that organizes all work around small features with visible value for the customer, each deliverable within a few days. It was designed in the late 1990s by Jeff De Luca, with Peter Coad, on a large banking project.
The method follows five steps: build an overall domain model, draw up the list of features (phrased like "calculate the total of an order"), plan by feature, design by feature, then build by feature. The last two repeat in a short loop.
FDD belongs to the agile family, but it is more structured than Scrum: it emphasizes upfront modeling, code ownership by developer and precise progress tracking. It is rarely practiced as such today, but its central idea, splitting work into customer-valued features, can be found in most product teams.
Feature Flag
MethodologyTerm= an on/off switch for a feature
A feature flag (or feature toggle) is a switch in the code that turns a feature on or off in production through simple configuration, without redeploying the application.
The new feature's code is shipped but stays hidden behind a condition. It is then opened to whoever you choose: the internal team, 5% of users, a specific customer, then everyone. If something goes wrong, it is switched off in seconds (kill switch).
Common uses: progressive rollouts, A/B testing, access limited to certain plans, and above all separating the moment code is deployed from the moment the feature is released. This is what makes trunk-based development possible, where everyone merges quickly into the main branch.
Tools such as LaunchDarkly, Unleash, GrowthBook or Flagsmith manage them. The downside: forgotten flags pile up and complicate the code if they are not cleaned up.
Feature Store
TechTerm= the pantry of ML models
A feature store is a system that centralizes, stores and serves the features (variables computed from raw data) used by machine learning models, both for training and for predictions in production.
A feature is, for example, “number of purchases by a customer over 30 days”. Without a feature store, each team recomputes it its own way, and the computation used for training often differs from the one used in production. The model then sees different data from what it learned on: this is training/serving skew, a classic source of silent degradation.
The feature store solves this with a single definition of each feature, historical (offline) storage for training and fast (online) storage to serve predictions within milliseconds.
Common solutions are Feast (open source), Hopsworks, Tecton, and the built-in feature stores of Databricks and Google Vertex AI. The tool is mostly justified when several models run in production.
Fine-tuning
TermMethodology= specializing an existing model
Fine-tuning means continuing the training of an already pre-trained AI model, such as an LLM, on a targeted dataset in order to specialize it for a task, a domain or a style without starting from scratch.
The most common form is supervised fine-tuning: you supply examples of questions and expected answers. Other methods adjust the model based on human preferences (RLHF, DPO). To reduce costs, techniques such as LoRA or QLoRA train only a small portion of the parameters, which makes the operation feasible on a few GPUs.
Fine-tuning is used to enforce an output format, a tone or domain vocabulary, or to get a small model to match a large one on a specific task, in order to lower inference costs.
It is not a good way to add knowledge that changes often: for that, RAG or a better prompt are usually simpler and easier to update. The choice depends on the problem, and is validated through evaluation.
FinOps
MethodologyConcept= keeping the cloud bill in check
FinOps (a blend of Finance and DevOps) is a cloud cost management practice that makes technical teams accountable for what they consume, working with finance and leadership.
The principle: make spending visible and attributable. Resources are tagged by team and product, costs are tracked continuously, then action follows: deleting forgotten resources, resizing oversized machines, shutting down test environments at night, committing to volumes to get discounts (reserved instances, savings plans).
FinOps is not just cost-cutting. It is about balancing cost, performance and reliability, and knowing how much a customer or a feature costs. The topic is growing with AI, whose GPU compute and model call costs can climb fast.
The FinOps Foundation, part of the Linux Foundation, publishes a reference framework and certifications.
Fork
Term= the copy that breaks away
A fork is an independent copy of a software project, made from its source code and evolved separately, without depending on the original project's choices.
On GitHub or GitLab, forking a repository is an everyday move: you create your own copy, make your changes there, then propose merging them back into the original project through a pull request (or merge request). It is the basic mechanism of open-source contribution.
The word also has a stronger meaning: a split. When part of the community disagrees with a project's direction (governance, license, acquisition), it forks it and carries it on under another name. Well-known examples:
- LibreOffice, from OpenOffice
- MariaDB, from MySQL
- OpenSearch, from ElasticsearchNot to be confused with a clone, which is simply a local copy of a repository to work on.
Forward Deployed Engineer / FDE
Role= the engineer embedded at the client
The Forward Deployed Engineer (FDE) is a software engineer sent to the client to connect their company's platform to that client's real-world uses: data, existing systems, business processes.
The role was popularized by Palantir, then adopted by many publishers of complex software and AI. The FDE really writes code: connectors, data pipelines, custom applications, prompts and agents. But they do so in direct contact with the client's teams, whose business and constraints they understand. They bring what they learn in the field back to the product.
It is neither a consultant making slides, nor an IT services developer placed on contract, nor a presales engineer. It is a product engineer working as close as possible to the value created for the client.
FPGA
Tech= the chip you rewire with software
An FPGA (Field-Programmable Gate Array) is an electronic circuit whose internal logic can be reconfigured after manufacturing, which makes it possible to create a custom component without fabricating a dedicated chip.
Instead of executing a program instruction by instruction like a processor, an FPGA is "wired" to perform a function directly, with very high parallelism and very low, predictable latency. It is described in a hardware description language such as VHDL or Verilog, then tools translate it into a circuit configuration. The main manufacturers are AMD (formerly Xilinx) and Altera.
FPGAs are found in radar signal processing, telecoms, electronic warfare, imaging, high-frequency trading and chip prototyping. It is a key technology in defense and space.
Framework
TermTech= the skeleton ready to build on
A framework is a structured set of components, rules and conventions that serves as the foundation for building an application, so you do not start from scratch on every project.
It provides the common building blocks (routing, data access, security, testing) and imposes a way of organizing code. The difference with a plain library: with a library, your code calls it when it needs to; with a framework, the framework structures the application and calls your code at the right places.
Each framework is tied to a language. A few examples:
- PHP: Symfony, Laravel
- Java: Spring
- Python: Django, FastAPI
- JavaScript / TypeScript: Angular, Next.js, NestJSChoosing a framework commits a team for years: it shapes how code is written, which ecosystem is available and which pool of developers you can hire from.
From scratch
Term= starting from a blank page
Building software from scratch means starting from zero, without reusing existing code, rather than evolving an application already in place.
The phrase comes up in two situations. First: a startup writing its first product, which means making every founding choice (language, architecture, hosting). Second: a complete rebuild of an existing system that has become too costly to maintain.
Starting from zero gives freedom and avoids accumulated technical debt. But it is a risky bet on a product already in production: you have to reproduce years of often poorly documented business rules, and keep the old system running in parallel. Many teams prefer a gradual migration, replacing the system piece by piece.
From scratch does not mean without tools: you still rely on frameworks, libraries and cloud services.
Front-end
TermConcept= the part the user sees
The front-end is the part of an application users interact with directly: the interface displayed in the browser or on screen, as opposed to the back-end running on servers.
A front-end developer turns designers' mockups into a working interface. They rely on three building blocks: HTML for structure, CSS for styling, JavaScript (often TypeScript) for interactivity. Most projects use a framework or library such as React, Vue.js, Angular or Svelte.
The job goes well beyond integration work:
- managing application state and API calls
- loading and rendering performance
- accessibility and cross-browser compatibility
- adapting to every screen size (responsive)The boundaries are shifting: with server-side rendering (Next.js, Server Components), part of front-end work now runs on the server.
Fullstack
Role= the developer on both sides
A fullstack developer is a developer able to work on both the front-end (the interface) and the back-end (server, API, database) of an application.
In practice, they can ship a feature end to end: the screen, the business logic, the data storage. The most common fullstack stacks pair a front-end framework (React, Vue.js) with a back-end in Node.js, PHP, Python or Java, sometimes with TypeScript on both sides.
In reality, most fullstack developers have a dominant side: they are back-end with some front-end, or the reverse. Being truly expert in both remains rare, especially since the role often includes some infrastructure and deployment work.
The profile is in high demand in young companies, where a small team has to cover everything. As the team grows, roles become more specialized.
Gemini
Tech= Google's multimodal AI
Gemini is the family of AI models developed by Google DeepMind, introduced in late 2023 to succeed the PaLM models. The name also refers to Google's consumer assistant, which replaced Bard in 2024.
Gemini was designed from the start as multimodal: a single model handles text, images, audio, video and code. It comes in several sizes, from lightweight versions for fast use or on-device use on phones to the most capable versions. Its models stand out for very large context windows, exceeding a million tokens.
It is available through the Gemini app, the Gemini API via Google AI Studio and the Vertex AI platform on Google Cloud. Google also builds it into its own products: search, Gmail, Docs and the rest of Google Workspace, Android.
Google also releases Gemma, a family of smaller open-weights models stemming from the same work.
GitOps
MethodologyTech= Git as production's remote control
GitOps is a way of operating infrastructure and deployments in which a Git repository describes the desired state of the system, and an automated tool makes sure reality matches what is written.
In practice, application configuration (often Kubernetes manifests or Helm charts) lives in Git. An agent installed in the cluster, such as Argo CD or Flux, continuously compares the actual state with the declared state and corrects any drift. To deploy a new version, you do not run a command on the server: you merge a pull request.
The benefits come from Git itself: a history of every change, review by a colleague before it is applied, rollback by reverting a commit, and rebuilding an identical environment.
The term was coined by Weaveworks in 2017. GitOps is mostly used with Kubernetes, alongside Infrastructure as Code (Terraform, OpenTofu), which creates the underlying infrastructure.
Go
LanguageTech= the simple language of cloud infra
Go (or Golang) is a compiled programming language designed at Google by Robert Griesemer, Rob Pike and Ken Thompson, made public in 2009, and built to write fast, reliable server software simply.
Its philosophy: few keywords, a deliberately minimalist syntax, very fast compilation that produces a single binary that is easy to deploy. Its main strength is concurrency: goroutines and channels make it possible to handle thousands of parallel tasks without the usual complexity.
Go has become the language of cloud infrastructure. Docker, Kubernetes, Terraform and Prometheus are written in Go. It is also found in high-traffic back-ends, microservices, APIs and command-line tools.
Its simplicity is a deliberate choice: Go favors code the whole team can read over expressiveness.
GPT
Tech= the engine behind ChatGPT
GPT (Generative Pre-trained Transformer) is the family of large language models developed by OpenAI since 2018. It brought generative AI to the general public with the launch of ChatGPT in November 2022.
The name describes the method. The model is pre-trained on huge volumes of text to predict the next word (more precisely the next token), using the Transformer architecture. It is then adjusted, notably through RLHF, to follow instructions and comply with safety rules.
Generations have followed one another (GPT-3, GPT-4, GPT-5…), with versions able to handle images and audio, and so-called reasoning models that take more time to answer complex problems. They are available through ChatGPT, the OpenAI API and Microsoft Azure.
In everyday language, “GPT” is often used to mean any LLM, which is a misuse: Claude, Gemini or Mistral rely on similar principles but are not GPTs.
GPU
TermTech= the chip that runs AI
A GPU (Graphics Processing Unit) is a processor designed to run thousands of simple calculations in parallel. Originally built for graphics rendering, it has become the reference chip for training and running AI models.
Where a CPU handles a few complex tasks very quickly, a GPU applies the same operation to huge arrays of numbers at once. That is exactly what neural networks need: matrix multiplications at very large scale. The memory on the card (VRAM) matters as much as raw power, because it limits the size of the model you can load.
Nvidia dominates this market thanks to its chips and above all CUDA, its software environment. AMD, Google (with its TPUs) and others offer alternatives. GPUs are rented from hyperscalers or specialized clouds, or bought for in-house servers.
GraphQL
LanguageTech= an API à la carte
GraphQL is a query language for APIs, together with a server-side engine, created by Facebook in 2012 and released as open source in 2015. It lets the client request exactly the data it needs, in a single query.
With a classic REST API, you often call several endpoints and receive fields you do not need. With GraphQL, the server exposes a typed schema describing all available data, and the client writes a query specifying the fields it wants. There are three types of operations: queries to read, mutations to write, subscriptions to receive real-time updates.
It is useful when several interfaces (web, mobile, partners) consume the same data with different needs. In return, caching, security (overly heavy queries) and server-side performance require more work than with REST.
The ecosystem includes Apollo, Relay and urql on the client side, and Apollo Server, GraphQL Yoga or Hasura on the server side. The specification is maintained by the GraphQL Foundation.
Guardrails
TermConcept= the safety rails around AI
Guardrails are the set of controls placed around an AI model in production to govern what it receives, what it produces and what it is allowed to do.
They work at several levels. On input: detecting prompt injection attempts, filtering prohibited content or personal data. On output: checking the format, blocking toxic content or leaks of sensitive information, making sure the answer stays within the intended scope. For an agent: limiting the tools it can access, requiring human approval before an irreversible action.
They are necessary because an LLM is not deterministic: the same request can produce different answers, and a malicious user may try to hijack it.
They are implemented with coded rules, dedicated classification models (such as Llama Guard), libraries (Guardrails AI, NeMo Guardrails) or the moderation services of model providers and clouds.
Hallucination (AI)
TermConcept= a confident mistake
A hallucination is a false or made-up answer that a generative AI model presents with the same confidence as accurate information: a quote that does not exist, fictitious case law, a code function that never existed.
The phenomenon comes from the very way LLMs work. They produce the most plausible text, token after token, based on what they have learned, without checking facts against a source. The risk rises on niche or rare topics, precise figures, references, and when the question assumes an answer that does not exist.
It cannot be eliminated entirely, but it can be reduced:
- grounding answers in reliable documents with RAG, and having sources cited;
- explicitly allowing the model to answer that it does not know;
- checking outputs with rules, a second model or a human;
- continuously measuring the error rate on a test set.Helm
ToolTech= the package manager for Kubernetes
Helm is the package manager for Kubernetes: it lets you package, share and install a complete application, with all its Kubernetes resources, as a package called a chart.
Deploying an application on Kubernetes often takes a dozen configuration files (deployments, services, configurations, secrets…). A chart bundles them as parameterized templates: a values file (values.yaml) adapts the same chart to each environment, for example more replicas in production than in testing.
Helm also tracks installed versions, upgrades, rollbacks and dependencies between charts. Most software in the Kubernetes ecosystem (Prometheus, Argo CD, cert-manager…) is distributed as a chart.
Helm is a graduated project of the Cloud Native Computing Foundation. It is often used with a GitOps tool such as Argo CD or Flux, or as an alternative to Kustomize.
HTML
LanguageTech= the structure of a web page
HTML (HyperText Markup Language) is the markup language that describes the structure and content of a web page: headings, paragraphs, links, images, forms.
It is not a programming language: it contains no logic. It tells the browser what each element is, using tags (h1 for a heading, p for a paragraph, a for a link). Styling is handled by CSS, interactivity by JavaScript.
Created by Tim Berners-Lee in the early 1990s along with the web itself, HTML is now a living standard, updated continuously rather than through major versions.
Well-written HTML is called semantic: each tag matches the meaning of the content. This matters for accessibility (screen readers), for search engine optimization and for code maintainability.
HTTP / HTTPS
Term= the common language of the web
HTTP (HyperText Transfer Protocol) is the protocol that lets a client, usually a browser, request a resource from a server and receive its response; HTTPS is its encrypted, secure version.
The principle is simple: the client sends a request (a method such as GET to read or POST to send data, an address, headers), and the server replies with a status code (200 if all is well, 404 if the page does not exist, 500 for a server error) and content. HTTP keeps no memory from one request to the next: cookies and tokens are used to maintain a session.
HTTPS adds an encryption layer (TLS) that protects exchanges against eavesdropping and tampering. It has become the norm on the web.
HTTP is not only for web pages: most APIs are built on it. The protocol has evolved with HTTP/2 and then HTTP/3 to become faster.
Hugging Face
Tool= the GitHub of AI models
Hugging Face is a company and an online platform where the AI community shares models, datasets and demos, much like GitHub for code.
The platform (the "Hub") hosts a very large number of open-weight models, downloadable in a few lines of code. The company also develops widely used open-source libraries, including Transformers, which loads and uses these models in a standard way, as well as training and inference tools such as TGI.
Founded by French entrepreneurs, Hugging Face has become an almost mandatory stop for anyone working with open models: people publish, compare and retrieve models there. It also offers paid plans to host and serve models.
Hyperscaler
Term= the global cloud giant
A hyperscaler is a cloud provider able to operate infrastructure at very large scale, in datacenters spread across several continents. The term mainly refers to AWS (Amazon), Google Cloud and Microsoft Azure.
These players offer hundreds of services: compute, storage, databases, networking, AI, security. Their size lets them add capacity very quickly and provide a global offering, which few players can do. They are also among the main providers of rented GPUs for AI.
In Europe, the word often comes up in the sovereignty debate: the three main hyperscalers are American and therefore subject to US law, including the Cloud Act. This pushes some sensitive organizations toward sovereign cloud offerings or hybrid partnerships.
IaaS
Term= servers rented on demand
IaaS (Infrastructure as a Service) is a cloud computing model in which a company rents basic computing resources (virtual machines, storage, networking) from a provider instead of buying and maintaining its own servers.
The provider manages the physical hardware, the datacenters and the virtualization. The customer keeps control of everything else: operating system, configuration, middleware, applications, data security. They pay for what they use and can scale resources up or down within minutes.
It is the cloud layer that leaves the most control, and therefore the most responsibility. Above it come PaaS (the provider also manages the runtime environment) and SaaS (ready-to-use software).
The best-known IaaS offerings come from AWS, Microsoft Azure and Google Cloud, and in France from OVHcloud and Scaleway. They are increasingly managed through code (Terraform).
IaC
MethodologyTech= infrastructure written as code
IaC (Infrastructure as Code) is a practice of describing IT infrastructure (networks, servers, databases, access rights) in versioned code files, instead of configuring it by hand in a console.
These files are stored in Git, reviewed like application code, then applied by a tool. The gains are reproducibility (recreating an identical environment), traceability (who changed what and when), review before changes, and the ability to spin up temporary environments on demand.
Two main approaches coexist. The declarative approach describes the desired state and lets the tool work out the changes: Terraform or its open source fork OpenTofu, Pulumi, AWS CloudFormation. The more procedural approach runs configuration steps on existing machines: Ansible is the most widespread example.
IaC is a pillar of DevOps and the foundation of GitOps.
Inference
TermConcept= using the model
Inference is the stage where an already trained AI model is used to produce a result from new data: a prediction, a classification, an LLM answer. It stands in contrast to training, which is used to build the model.
For an LLM, every request is an inference: the model receives a prompt and generates its answer token by token. Three indicators matter: latency (the time to the first token), throughput (tokens produced per second) and cost, which depends on the hardware, usually GPUs.
Training is a one-off effort, while inference repeats with every use. At scale, it often becomes the largest cost item of an AI product.
Hence an engineering field dedicated to optimizing it: quantization, grouping requests (batching), caching, speculative decoding. It relies on specialized engines such as vLLM, TensorRT-LLM or llama.cpp, and on inference platforms that rent out this service.
Inference platform
TechTool= the engine that serves the model
An inference platform is the software and hardware infrastructure that runs an AI model in production and answers user requests quickly, reliably and at a controlled cost.
Training a model is not enough: it then has to be loaded onto GPUs, handle thousands of concurrent requests, batch requests to make better use of hardware, monitor latency and absorb load spikes. An inference platform combines for this an execution engine (such as vLLM), orchestration (often Kubernetes), an API and observability.
It can be built in-house or bought from a specialized provider that bills by usage. The choice depends on volume, confidentiality constraints and the team's ability to operate the whole stack.
Information System
Term= the company's nervous system
The information system (IS) is the set of human, organizational and technical means that allow an organization to collect, store, process and share the information it needs to operate. It is not limited to IT: it also includes the processes and people who produce and use information.
Its technical side covers hardware, networks, software (ERP, CRM, business tools, email) and data. The department in charge of it is the IT department, led by a CIO.
In a large group, the IS is often a heterogeneous stack built over decades, with legacy applications that must coexist with the cloud and newer tools. Its security, its overall architecture and its cost are executive-level topics.
Integration
Concept= making software talk together
Integration is the work of connecting different applications, data and systems so that they function together as a coherent whole.
In a company, tools multiply: CRM, ERP, billing tool, business application, data warehouse. Without integration, information is re-entered by hand and drifts apart. Integration relies on APIs, scheduled data flows, message buses (such as Kafka) or dedicated platforms.
The word has other meanings worth telling apart:
- web integration: turning a graphic mockup into HTML and CSS pages, the job of an HTML integrator
- continuous integration: automatically merging and testing code on every change (the CI in CI/CD)
- system integration, in industry and defense: assembling hardware and software into a working productIntegration is often where the costliest problems of an information system hide.
IoT
TermConcept= objects that send their data
IoT (Internet of Things) refers to physical objects equipped with sensors, electronics and a network connection, able to collect data and transmit it to other systems.
A complete IoT system spans several layers:
- the device: sensors, microcontroller, embedded software, power management
- connectivity: Wi-Fi, Bluetooth Low Energy, cellular networks, LoRaWAN
- the platform: receiving, storing and processing data, often in the cloud
- the application: dashboards, alerts, remote controlIt shows up in the connected home, but above all in industry (predictive maintenance, machine monitoring), energy (smart meters), healthcare (monitoring devices), logistics and agriculture.
IoT's specific challenges are securing large numbers of poorly protected devices, power consumption, remote updates and data volume.
IP Address
Term= a machine's postal address on the network
An IP address (Internet Protocol) is the numerical identifier that lets a device send and receive data on a network using the IP protocol, starting with the internet. Without it, a data packet would not know where to go.
There are two versions. IPv4 is written as four numbers separated by dots (192.168.1.10) and allows only about 4.3 billion addresses, a pool exhausted long ago. IPv6, much longer, offers a virtually unlimited space and is being rolled out gradually.
An address is not always unique or fixed: corporate networks and home routers use private addresses, shared behind a single public address (NAT), and many addresses are assigned dynamically (DHCP). The concept shows up everywhere: network configuration, cloud, firewalls, cybersecurity, rough geolocation of a visitor.
IT infrastructure
TermConcept= the foundations that run the software
IT infrastructure is the set of hardware, network and software resources that allows an organization to run its applications and digital services.
It is made up of several layers:
- hardware: servers, storage, workstations, datacenters
- network: routers, firewalls, connectivity, DNS
- core software: operating systems, virtualization, databases, monitoring toolsIt can be hosted on premises, in the cloud or in a hybrid setup. The underlying trend is to manage it as code (infrastructure as code): resources are described in versioned files, then tools create and configure them automatically.
Well-designed infrastructure is judged on availability, security, ability to absorb growth and cost. When it works, nobody notices it; when it goes down, the whole business stops.
ITAR
TermConcept= US control over military exports
ITAR (International Traffic in Arms Regulations) is the US regulation that controls the export, transfer of and access to defense articles, technologies and services listed on the United States Munitions List (USML).
Its effect goes beyond US borders: equipment made in Europe that includes an ITAR-controlled component cannot be resold or re-exported without US authorities' approval. Technical data is also covered, which restricts some people's access to information based on their nationality, unless authorized.
For European defense and space manufacturers, ITAR weighs on supplier choices and on the freedom to export. Hence the search for so-called "ITAR-free" products, designed without any component subject to this regulation, often presented as a sovereignty argument.
Java
LanguageTech= the language of large enterprise systems
Java is an object-oriented programming language created by Sun Microsystems and released in 1995, designed so that the same program runs on any system thanks to a virtual machine, the JVM.
Java code is compiled into an intermediate format (bytecode), which the JVM runs on Windows, Linux or macOS. Sun was acquired by Oracle in 2010, which has steered the language since, with a new release every six months. Open distributions such as OpenJDK are widely used.
Java underpins a great many critical systems: banks, insurers, telecoms, large e-commerce sites, public administrations. Its ecosystem is massive, with Spring as the dominant server-side framework. The JVM also runs other languages, such as Kotlin (widely used on Android) and Scala.
Its strengths: robustness, performance, mature tooling, and a huge pool of developers.
JavaScript
LanguageTech= the language of the interactive web
JavaScript is a programming language created in 1995 by Brendan Eich at Netscape, which became the only language run natively by every web browser, making it the language of interactivity on the web.
Together with HTML (structure) and CSS (styling), it forms the core trio of front-end development: it responds to clicks, updates the page without reloading it, and talks to servers. It is standardized as ECMAScript and evolves every year.
Since Node.js, JavaScript also runs on the server, making it possible to use a single language across an entire application. Its ecosystem is immense: React, Vue.js, Angular for the interface, Express or NestJS for the back-end, and millions of packages on npm.
Despite its name, it has nothing to do with Java. Many teams now use it through TypeScript, which adds static typing.
Kafka
TechTool= the conveyor belt for events
Apache Kafka is an open source distributed event streaming platform, born at LinkedIn, open-sourced in 2011 and later made an Apache Software Foundation project.
Kafka records message streams in topics split into partitions, stored on disk and replicated across several servers. Order is guaranteed within a partition. Producers and consumers do not know each other: each one reads at its own pace, and a consumer can replay the history.
It is used to move data in real time between systems: change data capture from a database (CDC), event-driven microservice architectures, streaming analytics or ML pipelines. Confluent, founded by Kafka's creators, is its main commercial vendor; AWS (MSK), Aiven and others offer managed services.
Kanban
Methodology= the board that limits ongoing work
Kanban is a work management method that makes the flow of tasks visible on a board and limits how many tasks are carried out in parallel, in order to deliver smoothly and continuously.
It comes from manufacturing: Toyota developed it after the Second World War, driven by engineer Taiichi Ohno, to produce at the pace of actual demand and avoid unnecessary inventory. It was adapted to software development in the 2000s.
The board is split into columns representing the stages (to do, in progress, in review, done). The key principle, often forgotten, is the work-in-progress limit (WIP limit): you do not start a new task while the column is full. This forces teams to finish before starting and exposes bottlenecks.
Unlike Scrum, Kanban imposes neither sprints nor dedicated roles. It suits teams with an unpredictable flow, such as support, operations or maintenance.
Kubernetes
TechTool= the conductor for containers
Kubernetes (abbreviated k8s) is an open source container orchestrator, launched by Google in 2014 based on its internal experience with Borg, and handed over to the Cloud Native Computing Foundation (CNCF) in 2015.
It automatically deploys, scales and restarts containerized applications across a set of machines. You describe the desired state (how many instances, which resources, which services are exposed) in configuration files, and the control plane continuously brings the actual state back to that desired state.
It is the de facto standard for cloud-native deployment. All major clouds offer it as a managed service (EKS at AWS, GKE at Google, AKS at Microsoft), as do European players such as OVHcloud, Scaleway and Clever Cloud.
Laravel
FrameworkTech= the PHP framework built for speed
Laravel is an open-source framework written in PHP, created by Taylor Otwell in 2011 and released under the MIT license, used to build web applications and APIs with a syntax known for being clear and pleasant to use.
It follows the MVC pattern and provides out of the box what an application needs: routing, authentication, queues, email sending, scheduled tasks. Its ORM, Eloquent, lets developers handle the database as objects, and its command-line tool, Artisan, automates repetitive tasks.
Laravel relies on a rich ecosystem maintained around the project: deployment tools, admin panel, payment handling, reactive interfaces. It is one of the two major PHP frameworks alongside Symfony, several of whose components it actually reuses.
It is often found in startups, agencies and SaaS products that want to ship quickly.
Layer 2
TechTerm= the blockchain's fast lane
A Layer 2 (or L2) is a network built on top of a main blockchain (the Layer 1, such as Ethereum) to process transactions faster and more cheaply, while relying on that blockchain's security.
The most common form is the rollup: it executes a large number of transactions off the main chain, then publishes a summary to it. Optimistic rollups (Arbitrum, Optimism, Base) treat transactions as valid unless challenged; zk-rollups (zkSync, Starknet, Linea) publish a cryptographic proof of validity.
L2s make applications usable that Layer 1 fees would make too expensive: DeFi, payments, games, NFTs.
Lead Dev / Lead Tech
Role= the developer who guides the team
A Lead Dev (or Lead Tech) is an experienced developer who, in addition to coding, carries a team's technical responsibility: code quality, architecture choices, and helping other developers grow.
Their daily work combines three dimensions:
- technical: they still code a significant share of their time and make the structural calls
- quality: code reviews, standards, testing, reducing technical debt
- mentoring: supporting junior profiles, working with product to turn needs into solutionsThe role varies a lot between companies. In a small team, the Lead Dev is often the technical reference for the whole product. In a larger organization, they lead one team among others, under an Engineering Manager or a CTO.
It differs from the Engineering Manager, focused on people and organization, and from the Staff Engineer, who has cross-team technical influence without a dedicated team.
LLM
TechTerm= the engine behind writing AIs
An LLM (Large Language Model) is an artificial intelligence model trained on very large volumes of text to predict the next token in a sequence, which allows it to write, summarize, translate, reason and write code.
LLMs rely on the transformer architecture and range from a few billion to several hundred billion parameters. They are first pre-trained on raw text, then adjusted (instruction tuning, RLHF) to follow instructions and respect behavioral rules.
The best-known families are Claude (Anthropic), GPT (OpenAI), Gemini (Google), Mistral, Llama (Meta) and Qwen (Alibaba). They are used through an API or, for open-weight models, on one's own infrastructure.
LLMOps
MethodologyConcept= running AI applications in production
LLMOps is the set of practices for putting LLM-based applications into production and running them: versioning prompts like code, continuously evaluating answers, tracing model calls, controlling token costs and detecting drift.
It inherits from MLOps but changes focus: you rarely train the model yourself, you orchestrate prompts, tools, document retrieval (RAG) and guardrails around an external model. Typical tools cover evaluation (Promptfoo, Braintrust, LangSmith) and observability (Langfuse, Helicone, Arize).
It is what separates a prototype that impresses in a demo from a reliable service: without automated tests or monitoring, a prompt or model update can degrade the product without anyone noticing.
Machine Learning
TermConcept= learning from examples
Machine learning is a branch of artificial intelligence in which a program learns to perform a task from data, instead of following rules written by hand by a developer.
Instead of coding if the message contains this word, it is spam, you give the model thousands of already-labeled examples and it works out the patterns on its own. There are three main families:
- supervised learning: learning from labeled examples (predicting a price, detecting fraud)
- unsupervised learning: finding groups or structures in raw data
- reinforcement learning: learning through trial and rewardDeep learning, which uses deep neural networks, is a subset of machine learning; it is what made LLMs possible. But many real-world problems (scoring, forecasting, recommendation) are still solved very well with more classic methods, which are simpler to explain and maintain.
Marketplace
Term= the platform between sellers and buyers
A marketplace is a digital platform that connects independent sellers and buyers, and generally earns a commission on each transaction, without owning the products or services being sold.
The model comes in several forms: business to consumer (Amazon for its third-party seller side), consumer to consumer (Vinted), for services (Airbnb, freelance platforms) or between businesses (B2B marketplaces for supplies or industrial parts).
Its main difficulty is the chicken-and-egg problem: no sellers means no buyers, and vice versa. A marketplace has to kick-start both sides at once, then maintain trust (reviews, secure payment, dispute resolution).
Technically, a marketplace combines demanding building blocks: search engine, payments with payouts to sellers, account management and fraud prevention.
MCP
TechTerm= the universal plug for AI agents
MCP (Model Context Protocol) is an open protocol, published by Anthropic in November 2024, that standardizes how an AI application (assistant, agent, code editor) connects to external tools and data sources.
An MCP server exposes three kinds of elements: resources to read, tools to call (actions) and prompt templates. An MCP client, such as Claude, Cursor or VS Code, uses them all in the same way. An integration written once therefore works with several AI applications.
The protocol has been adopted beyond Anthropic, notably by OpenAI and Google. Many vendors publish their own MCP server (GitHub, Notion, Slack, Stripe…) so that agents can act inside their products.
MFA / SSO
TermConcept= two keys, one door
MFA (Multi-Factor Authentication) is a mechanism that requires at least two proofs of identity of different kinds: something you know (password), something you have (phone, security key) or something you are (biometrics).
SSO (Single Sign-On) lets a user sign in once with an identity provider (Microsoft Entra ID, Okta, Google Workspace) and then access every connected application. It relies on protocols such as SAML 2.0 or OpenID Connect.
The two go together: SSO concentrates access in one place, MFA protects that place. Passkeys (the FIDO2 and WebAuthn standards) are gradually replacing passwords and one-time codes with passwordless sign-in that is more resistant to phishing.
Microservice
MethodologyTerm= the application split into small services
Microservices are a software architecture in which an application is split into small independent services, each responsible for a specific business function and deployable separately from the others.
Instead of one big program (a monolith), you get for instance one service for payments, one for the catalog, one for notifications. Each often has its own database and communicates with the others through APIs or messages. A team can change and deploy its service without waiting for the others.
The benefits: more autonomous teams, targeted scaling of the busiest parts, technical choices suited to each service. The cost is real:
- a distributed system, harder to test and debug
- a strong need for observability, automation and tools like Docker and Kubernetes
- distributed data that must be kept consistentMany teams go back to a modular monolith when the size of the organization does not justify that cost.
Mistral
Tech= generative AI made in France
Mistral AI is a French company founded in Paris in 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix, former researchers at Google DeepMind and Meta, that develops LLMs, some of which are released as open weights (weights you can download and run on your own infrastructure).
Its lineup covers general-purpose models (Mistral Large, Mistral Small), code (Codestral), vision (Pixtral) and reasoning (Magistral), as well as compact models built to run on modest hardware. The company also offers the Le Chat assistant and an API.
Mistral is often chosen by organizations that care about sovereignty or want to host their model themselves: government agencies, defense, industry, banks.
ML Engineer
Role= the one who ships models to production
The Machine Learning Engineer (ML Engineer) is the engineer who turns a machine learning model, often prototyped by a Data Scientist, into a reliable system running in production: data preparation, reproducible training, deployment, monitoring and updates.
They combine a Data Scientist's foundations (statistics, modeling) with a software engineer's (solid Python, testing, CI/CD, containers, cloud). They work with frameworks such as PyTorch or scikit-learn and tools such as MLflow, Weights & Biases, Vertex AI or SageMaker.
The role differs from the AI Engineer, focused on applications built around existing LLMs, and from the MLOps Engineer, focused on the platform and infrastructure that serve models.
MLOps
MethodologyConcept= DevOps for models
MLOps (Machine Learning Operations) is the set of practices that apply DevOps principles to the lifecycle of machine learning models: versioning code, data and models, automating training and deployment, monitoring quality and drift in production.
It addresses two concrete problems. Many models stay stuck at the prototype stage for lack of a deployment pipeline. And those that reach production degrade silently as real-world data drifts away from the training data.
Common tools are MLflow, Kubeflow, Vertex AI, SageMaker, Weights & Biases, Metaflow and DVC for data versioning. LLMOps is its counterpart for LLM-based applications.
MLOps Engineer
Role= the builder of the model factory
The MLOps Engineer is the engineer who builds and runs the platform that lets Data Scientists and ML Engineers ship their models continuously: training pipelines, model registry, deployment, monitoring, compute resource management.
The profile sits halfway between an SRE (reliability, monitoring, infrastructure as code, Kubernetes) and a data engineer (pipeline orchestration, large volumes, feature stores). They are responsible for the availability and performance of ML services in production.
They mostly work in organizations with several models in production or heavy compute needs (GPUs), where each team can no longer hack together its own pipeline.
Monorepo
MethodologyTerm= all the code in one repository
A monorepo is a single code repository that holds several distinct projects (applications, services, shared libraries), as opposed to a polyrepo where each project has its own repository.
The benefits: changing code used in several places in one go, easily sharing code and configuration, having an overall view. The trade-offs: a large repository and build tools that must understand internal dependencies so they only rebuild and retest what changed.
In JavaScript and TypeScript, teams use Turborepo, Nx or pnpm and Yarn workspaces. Bazel and Pants cover multi-language monorepos. Google and Meta are the best-known examples of very large-scale monorepos.
Multimodal
TermConcept= AI that sees, hears and reads
An AI model is called multimodal when it can understand or produce several types of data: text, images, audio, video, and sometimes other signals.
Rather than chaining specialized models (one to read the text in an image, another to analyze it), a multimodal model processes the different inputs in a single representation. You can ask a question about a photo, analyze a scanned document with its tables, or reply directly to a voice message.
The large models from Google, OpenAI, Anthropic and Mistral all accept at least text and images. Common uses: automated document reading, visual inspection in industry, voice assistants, video description.
MVC
Term= data, display and logic kept apart
MVC (Model-View-Controller) is a software architecture pattern that organizes an application into three distinct parts: the data, how it is displayed, and the logic linking the two.
Each part has a specific role:
- the model manages data and business rules (reading and writing to the database, validation)
- the view displays information to the user
- the controller receives user actions, calls on the model and picks the view to returnThis separation makes code more readable, easier to test, and allows the interface to change without touching business logic. The principle was born in the 1970s around the Smalltalk language.
It is the structure of many server-side web frameworks: Symfony, Laravel, Ruby on Rails, Spring MVC, or Django in a variant. Modern front-end interfaces have moved away from it toward component-based architectures.
MySQL
TechTool= the database of the classic web
MySQL is an open-source relational database management system (RDBMS), which stores data in linked tables and is queried with the SQL language.
Created in the mid-1990s by the Swedish company MySQL AB, co-founded by Michael Widenius (the name comes from his daughter's first name, My), it was acquired by Sun Microsystems in 2008, then came under Oracle's control when Oracle bought Sun. It is distributed under a dual license: open source (GPL) and commercial.
MySQL has historically been tied to the web: it is the M in the LAMP stack (Linux, Apache, MySQL, PHP) and WordPress's default database. It is easy to install, fast for reads and offered by every hosting provider and cloud.
After the Oracle acquisition, Michael Widenius launched a fork, MariaDB, which is largely compatible. In new projects, PostgreSQL is often preferred for its richer feature set.
Next.js
FrameworkTech= React ready for production
Next.js is an open source web framework built on React and maintained by Vercel, which adds what React needs to ship a complete site or application: routing, server-side rendering, static page generation and performance optimization.
In practice, each file in a folder becomes a page. The developer chooses page by page whether content is generated ahead of time, on each request or in the browser. Since version 13 and its App Router, Next.js is organized around React Server Components, which run part of the interface on the server.
It is one of the most common choices for starting a React project, from marketing sites to SaaS applications. It can be deployed on Vercel or on any hosting able to run Node.js.
NFT
TermTech= the digital certificate of ownership
An NFT (Non-Fungible Token) is a unique digital token recorded on a blockchain, used as a title of ownership or authenticity for a digital or physical object. Unlike a cryptocurrency, each NFT is distinct and cannot be swapped like-for-like with another.
On Ethereum, NFTs follow standards such as ERC-721 or ERC-1155, managed by smart contracts. The token usually contains a link to the content (image, file) rather than the content itself, which raises the question of how that content is preserved.
Popularized by digital art and collections such as CryptoPunks and Bored Ape, they went through a speculative bubble in 2021-2022, followed by a sharp decline. Remaining uses: ticketing, video game items, certificates, loyalty programs.
NIS2
TermConcept= Europe's cybersecurity obligation
The NIS2 directive (Network and Information Security 2) is the European text, adopted in 2022, that imposes cybersecurity obligations on so-called essential and important entities in 18 sectors: energy, transport, banking, health, water, digital infrastructure, cloud providers, managed service providers, public administration and manufacturing, among others. It replaces the 2016 NIS directive and greatly widens the number of organizations concerned.
It requires cyber governance owned by management, risk management measures (access control, MFA, vulnerability management, supplier security) and the reporting of serious incidents: early warning within 24 hours, notification within 72 hours. In France, ANSSI steers its implementation.
Fines can reach 10 million euros or 2% of worldwide turnover for essential entities.
Node.js
TechTool= JavaScript on the server
Node.js is an open-source runtime environment that runs JavaScript outside the browser, mainly on servers. It is neither a language nor a framework.
Created in 2009 by Ryan Dahl, it is built on V8, Google Chrome's JavaScript engine. Its distinctive feature is its asynchronous, non-blocking model: instead of waiting for a slow operation to finish (reading a file, calling a database), it moves on and handles the result when it arrives. This makes it efficient at handling many simultaneous connections.
It is mainly used to build APIs, real-time applications (chat, notifications), microservices and development tools. It ships with npm, the package manager that gives access to a huge ecosystem of libraries. The most common server-side frameworks are Express, Fastify and NestJS.
Its big advantage: a single language, JavaScript or TypeScript, across the whole product. Alternatives such as Deno and Bun follow the same principle.
Observability
MethodologyConcept= understanding why things break
Observability is the ability to understand what is happening inside a software system from the signals it emits: logs, metrics and traces, sometimes complemented by profiling and events.
Classic monitoring answers "is it working?" based on indicators defined in advance. Observability aims to answer "why is it behaving this way?", including for a failure nobody anticipated. It is essential in distributed architectures, where a single request crosses dozens of services.
Reference tools are Datadog, Grafana (with Prometheus, Loki and Tempo), Honeycomb, New Relic and Splunk. OpenTelemetry provides the standard instrumentation layer that feeds these tools.
OKR
Methodology= the goal and its measurable proof
OKRs (Objectives and Key Results) are a management method that sets, for a given period (often a quarter), a few qualitative objectives, each paired with two to five measurable key results that prove it has been reached. The method was formalized by Andy Grove at Intel, then brought to Google by John Doerr.
A good objective says where you want to go; a good key result measures an effect (conversion rate, response time, retention), not a completed task. Team OKRs connect to company OKRs and are visible to everyone.
The method is widespread in tech. Its classic failure: OKRs that simply list the backlog, or results always hit at 100%, a sign they lacked ambition.
Open source
Term= code open to everyone
Open-source software is software whose source code is published under a license that allows anyone to use, study, modify and redistribute it.
Open does not mean free of rules: the license sets the conditions. There are two main families:
- permissive licenses (MIT, Apache, BSD): the code can be reused almost freely, including in a closed commercial product
- so-called copyleft licenses (GPL): any modified version that is redistributed must stay under the same licenseOpen source does not mean free of charge or volunteer-run either. Many projects are backed by companies that earn a living from a model built around the code (hosting, support, advanced features), or by foundations.
Open source runs a large part of computing: Linux, Kubernetes, PostgreSQL, Python, React. For a company, it is a way to avoid vendor lock-in, but licenses and dependency security need monitoring.
Open weights
TermConcept= the freely downloadable model
An open-weights model is an AI model whose learned parameters, the "weights", are published and downloadable, which lets you run it on your own servers and adapt it.
Weights are the output of training: without them, you can only call the model through its publisher's API. With them, you can host it yourself, quantize it so it fits on more modest hardware, or specialize it through fine-tuning. Open weights does not always mean open source: training data and code often remain private, and the license may restrict some uses.
Llama (Meta), some Mistral AI models or Qwen (Alibaba) are well-known examples. These models are mostly distributed through Hugging Face. They appeal to organizations that want to keep their data in-house or control their costs.
OpenTelemetry
TechTool= the standard sensor for applications
OpenTelemetry (abbreviated OTel) is an open source application instrumentation standard, hosted by the CNCF, that provides common APIs, SDKs and tools to produce and collect traces, metrics and logs, whatever analysis tool is used afterwards. It was born in 2019 from the merger of the OpenTracing and OpenCensus projects.
Its value: you instrument the code once, then send the data to the tool of your choice (Datadog, Grafana, Honeycomb, New Relic…) without redoing everything if you switch. A central component, the collector, receives the signals, filters them and routes them to one or more destinations.
It is supported by most languages, frameworks and observability vendors, and serves as the foundation for cloud-native monitoring architectures.
Operating System / OS
TermTech= the machine's conductor
An operating system (OS) is the core software that manages a device's hardware resources (processor, memory, storage, peripherals) and gives applications a common framework to run in. Without it, every program would have to drive the hardware directly.
It shares computing time among programs, isolates their memory, manages files, networking and user permissions, and provides an interface (graphical or command line).
It is found everywhere:
- computers: Windows, macOS, Linux;
- smartphones: Android, iOS;
- servers and cloud: overwhelmingly Linux;
- embedded systems: real-time operating systems (RTOS) or embedded Linux, in automotive, aerospace, industry or defense.ORM
TermTool= the translator between code and database
An ORM (Object-Relational Mapping) is a tool that bridges an application's object-oriented code and a relational database, so data can be handled as objects without writing every SQL query by hand.
In practice, a customers table becomes a Customer class in the code. To save a customer, the developer creates an object and the ORM generates the matching SQL query. It also handles relationships between tables, schema changes (migrations), and protects against some vulnerabilities such as SQL injection.
Each ecosystem has its own:
- PHP: Doctrine (Symfony), Eloquent (Laravel)
- Java: Hibernate
- Python: SQLAlchemy, the Django ORM
- JavaScript / TypeScript: Prisma, TypeORMAn ORM saves a lot of time, but hides what happens in the database. Misused, it generates inefficient queries, such as the so-called N+1 problem, where a simple list triggers hundreds of queries.
PaaS
Term= turnkey hosting for applications
PaaS (Platform as a Service) is a cloud computing model in which a provider offers a ready-to-use platform to deploy and run applications, without the team having to manage servers, the operating system or the runtime environment.
The developer pushes their code and the platform handles the rest: installing dependencies, deployment, scaling, certificates, basic monitoring. The provider manages everything beneath the application; the customer remains responsible for their code and data.
PaaS sits between IaaS, where you manage your own virtual machines, and SaaS, where you use finished software. Examples: Heroku, Google App Engine, Render, and in France Clever Cloud and Scalingo.
Its strengths: starting fast and focusing on the product, without an operations team. Its limits: less control, costs that can climb with traffic, and a risk of vendor lock-in.
Pentest
MethodologyTerm= the commissioned, authorized break-in
A pentest (penetration test) is an offensive security audit in which an expert, with written authorization, attacks a target (web application, API, mobile app, network, cloud environment) to find exploitable flaws before an attacker does.
There are three approaches depending on the information provided: black box (no initial access), grey box (limited access, for example a user account) and white box (access to code and documentation). The deliverable is a report that ranks flaws by severity, with proof of exploitation and recommended fixes.
Pentesting complements automated tools (SAST, DAST), bug bounty programs and configuration audits. It is often required by enterprise customers and compliance processes.
PHP
LanguageTech= the historic language of websites
PHP is an open-source programming language designed for server-side web development: it generates pages and processes data before sending them to the browser.
Created in 1994 by Rasmus Lerdorf, it took off with the growth of the web thanks to its simplicity and availability at every hosting provider. It runs a considerable share of the web, including WordPress, Drupal and MediaWiki (the engine behind Wikipedia). Facebook was built in PHP in its early days.
Today's PHP has little in common with that of the 2000s: typing, mature object-oriented programming, much better performance. It is mostly used with frameworks such as Symfony and Laravel, and a dependency manager, Composer.
Its reputation is still shaped by old, unstructured code, whereas the modern language makes it possible to build robust applications.
Ping
TermTool= the network echo test
Ping is a network command that checks whether a machine can be reached on an IP network and measures how long a message takes to make the round trip.
It sends a small packet (an echo request from the ICMP protocol) to an IP address or a domain name, and waits for the reply. The result shows whether the machine responds, the round-trip time in milliseconds (latency) and the packet loss rate.
It is often the first reflex when diagnosing a connection problem: is the server online, is the network slow, are packets being lost along the way? No response does not prove the machine is down, since many firewalls deliberately block these requests.
By extension, to ping someone means to give them a quick nudge by message.
Platform Engineer
Role= the builder of developers' tools
The Platform Engineer is the engineer who builds and evolves an internal platform (Internal Developer Platform): a set of self-service tools that lets product teams deploy, monitor and secure their applications without depending on an infrastructure team.
Platform engineering extends DevOps. Rather than asking each team to set up its own tooling, a central team treats the platform as a product, with its users (developers), its roadmap and its metrics (deployment frequency, lead time to production, failure rate).
Day to day: Kubernetes, infrastructure as code (Terraform), CI/CD, secrets management, observability, and portals such as Backstage, created at Spotify.
Product
TermConcept= what the customer actually uses
In tech, the product is the software, app or service a company designs, puts in its users' hands and keeps improving over time. It is not a project delivered once and for all, but a living asset that is measured and refined.
By extension, "Product" also refers to the team and discipline that decide what to build: Product Managers, Product Owners, Product Designers. Their job is to understand needs, prioritize, measure usage and balance user value, technical feasibility and business interest.
A company is called "product-led" or "product-driven" when decisions start from real usage and measured outcomes, rather than from a frozen specification or a list of customer requests.
Product Manager
Role= the one deciding what to build
The Product Manager (PM) is the person responsible for what the team builds and why: they identify the user problems worth solving, set priorities and make sure the delivered product creates value for customers and for the company.
Their work is split between:
- discovery: user interviews, data analysis, competitive research
- prioritization: choosing what to do, and above all what not to do, in line with strategy
- delivery: working with developers and designers so the right features ship
- measurement: checking the real impact of what was deliveredThey usually have no line authority over the team: they persuade through the clarity of their vision and the quality of their arguments.
It differs from the Product Owner, a role defined by Scrum and more focused on managing the team's backlog, although the two are often confused in job ads.
Product Owner
Role= the keeper of the backlog
The Product Owner (PO) is the person who represents user needs to a development team: they decide what gets built first and make sure what ships actually meets the need. The role comes from Scrum, where it is one of three roles alongside the Scrum Master and the developers.
Day to day, they write and split user stories, order the backlog, define acceptance criteria and sign off on deliveries at the end of each sprint. They bridge users, business teams and the tech team.
The role is most common in large companies and organizations running agile at scale, where the PO often reports to a Product Manager who owns the product vision and strategy.
Prompt Engineer
Role= the instruction writer for AI
The Prompt Engineer is a profile that emerged with LLMs, responsible for designing, testing and improving prompts (the instructions given to generative AI models) to get reliable, safe and reproducible answers within a given product.
The work: writing system instructions, choosing the examples to give the model, structuring expected outputs, building test sets to compare versions and documenting what works. It combines an understanding of models, product sense, experimental rigor and writing quality.
The title still exists, but the role has largely merged into broader jobs: AI Engineer, AI Product Manager, domain specialists who design assistants. Dedicated positions are mostly found in AI labs and companies whose product relies entirely on a model.
Prompt Engineering
MethodologyTerm= the art of asking AI well
Prompt engineering is the set of techniques for writing and structuring the instructions given to a generative AI model, to get high-quality, reliable answers that match the expected behavior.
It relies on proven practices: giving examples (few-shot), asking the model to reason step by step, enforcing an output format (JSON for example), separating system instructions from user messages, providing the right context and planning for failure cases ("if the information is missing, say so").
With agents, the scope widens: describing the tools the model can call, managing what goes into the context window, chaining several steps. The quality of an AI application often depends as much on the instructions as on the chosen model.
Pure Player
Term= the online-only company
A pure player is a company whose entire commercial activity runs over the internet: it sells to and serves customers through a website or app, with no physical store or branch. The term is used in contrast to traditional players that added an online channel to an existing physical network.
A pure player may own warehouses, offices or data centers: what matters is that the customer relationship is fully digital. Early Amazon, Netflix, Spotify or an online-only newspaper such as Mediapart are classic examples.
The word is mostly used in retail, media and banking, sectors where digital disrupted the incumbents. Several pure players have since opened physical locations, which blurs the line.
PWA
TermTech= the website that installs like an app
A PWA (Progressive Web App) is a website that uses standard browser technologies to behave like a mobile app: installation on the home screen, offline operation, notifications, full-screen launch.
Two building blocks make it possible: a manifest, which describes the app (name, icon, colors), and a service worker, a script running in the background that caches content and receives notifications. Everything is served over HTTPS.
A PWA is distributed through a simple URL, without going through app stores, and a single codebase runs on every device. Android supports it well; Apple was long more restrictive, although iOS has accepted notifications from installed PWAs since 2023.
Python
Language= the Swiss Army knife of code
Python is a general-purpose programming language, created by Guido van Rossum and first released in 1991, known for its readable syntax and its rich ecosystem of libraries. It is interpreted: code runs without a manual compilation step, which makes it fast to write and test.
It covers very different uses:
- Data and AI: it is the dominant language for machine learning and data science (NumPy, pandas, PyTorch, scikit-learn).
- Web and back-end: with Django, Flask or FastAPI.
- Automation and scripting: tooling, system tasks, testing, engineering.Its simplicity has a cost: it is slower than compiled languages such as C++ or Rust, which is why heavy computation often relies on libraries written in C under the hood.
Quantization
MethodologyTech= the compressed model that runs cheaper
Quantization is a technique that reduces the precision of the numbers used to store an AI model's parameters, so that it takes less memory and runs faster.
A model stores its weights as floating-point numbers, often on 16 or 32 bits. Quantization brings them down to 8 bits, 4 bits or even less. The model becomes lighter, fits on a more modest GPU or a laptop, and responds faster. The trade-off: a possible loss of quality, which has to be measured for each use case.
There are several methods (GPTQ, AWQ, GGUF formats for local execution, etc.), applied after training or taken into account during it. Quantization is widely used to serve open-weight models in production and for embedded AI.
Quick & Dirty
Term= the patch that gets you by
A quick & dirty solution is one that is fast to put in place but not carefully built, chosen to fix or hide a problem right away rather than address it properly. It works, but at the cost of technical debt that will have to be paid back later.
It has its place in several situations: an emergency fix in production, a prototype or proof of concept (POC) to validate an idea, a customer demo. The risk appears when the patch stays in place for years and other components start relying on it.
A well-known anecdote: MS-DOS derives from QDOS (Quick and Dirty Operating System), a system written by Tim Paterson at Seattle Computer Products, which Microsoft bought in 1981.
RAG
MethodologyTech= AI that checks its sources
RAG (Retrieval-Augmented Generation) is an AI application architecture that, before answering, searches a document base for relevant passages and passes them to the LLM so it can rely on them. The term comes from a research paper published by Meta in 2020.
A typical RAG pipeline has three steps:
- ingestion: documents are split into pieces (chunks), turned into embeddings and stored in a vector database;
- retrieval: for each question, the closest pieces by meaning are found, often combined with keyword search;
- generation: the LLM writes its answer from these excerpts, ideally citing its sources.It is the standard method for making an AI answer on internal or recent data, and for limiting hallucinations. It is easier to update than fine-tuning: you just add documents.
ReactJS
FrameworkTech= the Lego bricks of the interface
React (also written React.js or ReactJS) is an open source JavaScript library created by Facebook (now Meta) and released in 2013, used to build user interfaces from reusable components. Each component manages its own state, and React updates the display automatically when that state changes.
Technically, it is not a full framework: React handles the view layer, and other tools are added for routing, data management or server-side rendering. Next.js is the most widely used framework for building complete applications on React. React Native applies the same principles to mobile apps.
It is one of the most widespread front-end technologies, from startups to large groups, which makes it a safe choice for hiring and maintaining a team.
Responsive
ConceptTerm= the page that fits the screen
A responsive page or app is an interface that automatically adapts its layout to the size of the screen displaying it, from smartphone to large monitor, without losing readability or ease of use. This is known as responsive web design.
In practice, the front-end developer uses CSS to rearrange elements based on available width: columns stacking on top of each other, a menu that collapses, resized images, adjusted text sizes. The key tools are flexible layouts (Flexbox, Grid) and media queries.
A large share of web traffic now comes from mobile, and many teams design for the small screen first (mobile first) before scaling up. Responsive is no longer an option but a baseline expectation.
RGPD / GDPR
TermConcept= Europe's personal data law
The GDPR (General Data Protection Regulation, RGPD in French) is the European regulation, applicable since May 25, 2018, that governs the processing of personal data of people located in the European Union, including by companies established outside the Union.
It gives rights to individuals (access, rectification, erasure, portability, objection) and imposes obligations on organizations: a legal basis for each processing activity, data protection by design (privacy by design), a record of processing activities, an impact assessment for high-risk processing, notification of breaches to the authority within 72 hours and, in some cases, the appointment of a data protection officer (DPO).
Fines can reach 20 million euros or 4% of worldwide turnover. In France, the supervisory authority is the CNIL.
RLHF
MethodologyTerm= teaching the model through human feedback
RLHF (Reinforcement Learning from Human Feedback) is an LLM training technique that, after pre-training, adjusts the model based on comparisons made by humans between several answers, so that it becomes more helpful, honest and safe.
The classic method has three steps: supervised fine-tuning on examples written by humans, training a reward model that learns to predict human preferences, then optimizing the LLM to maximize that score, usually with the PPO algorithm. Variants such as DPO learn directly from preferences, without a separate reward model.
RLHF played a key role in taking LLMs from research to the general public, starting with ChatGPT. Anthropic proposed Constitutional AI, where part of the feedback is produced by a model based on written principles (known as RLAIF).
Ruby
Language= the language built for developers
Ruby is an open source, interpreted, object-oriented programming language, created in Japan by Yukihiro "Matz" Matsumoto and released in 1995, with the explicit goal of making programming enjoyable for developers. Its syntax is concise and close to natural language.
Its popularity comes mostly from Ruby on Rails, a web framework released in 2004 that set strong conventions ("convention over configuration") and makes it possible to build a complete application very quickly. GitHub, Shopify and Airbnb were largely built with Rails.
Ruby remains very present in startups and scale-ups that chose it early, especially for web and SaaS applications. It is less used for data, AI or heavy computation.
SaaS
Term= software you rent online
SaaS (Software as a Service) is a model where software is hosted by the vendor in the cloud and used by the customer through a browser or app, in exchange for a subscription, instead of being installed on the customer's own machines.
For the customer, it removes installation, maintenance and updates, and makes it possible to get started fast. For the vendor, it creates recurring revenue and lets them ship improvements continuously to all customers at once. Slack, Salesforce, Notion and Zendesk are well-known examples.
SaaS differs from IaaS (renting infrastructure) and PaaS (renting a platform to deploy your own code). It places heavy demands on availability, data security and multi-customer management (multi-tenant).
SAFe
Methodology= agile, large-enterprise edition
SAFe (Scaled Agile Framework) is a proprietary methodology framework, created by Dean Leffingwell and marketed by the company Scaled Agile, that aims to apply agility to large organizations where dozens or even hundreds of teams work on the same products.
It groups teams into Agile Release Trains synchronized on cycles called Program Increments, often about a quarter long. Each cycle opens with PI Planning, a large shared planning event. SAFe also defines specific roles (Release Train Engineer, Product Management, System Architect) and several levels, up to the portfolio.
It is widespread in large corporations (banking, insurance, telecoms, industry, public sector). It is also criticized for its heaviness and its distance from the spirit of the Agile Manifesto; LeSS and the "Scrum of Scrums" are lighter alternatives.
SAST / DAST
MethodologyTool= the flaw detectors for code
SAST (Static Application Security Testing) and DAST (Dynamic Application Security Testing) are the two main families of automated application security testing: the first analyzes code without running it, the second attacks the application while it runs.
SAST reads the source code to spot known flaws (SQL injection, XSS, hard-coded secrets, weak cryptography). It fits into CI and flags problems before release. Tools: Semgrep, SonarQube, GitHub CodeQL, Snyk Code, Checkmarx.
DAST tests the deployed application from the outside, like an attacker: it sends crafted requests and observes the responses. It finds problems invisible in the code (configuration, authentication). Tools: OWASP ZAP, Burp Suite, Acunetix.
The two complement each other, along with dependency analysis (SCA), container scanning and infrastructure as code scanning.
SBOM
TermConcept= the ingredient list of a piece of software
An SBOM (Software Bill of Materials) is a structured, machine-readable inventory of every component that makes up a piece of software: libraries, direct and indirect dependencies, versions, licenses and hashes.
In practice, it is generated automatically at build time and stored with each shipped release. When a vulnerability is disclosed in a widely used library, teams query their SBOMs to find out within minutes which products are affected, instead of combing through every repository by hand.
The two standard formats are SPDX (backed by the Linux Foundation) and CycloneDX (backed by OWASP). SBOMs have been required to sell software to the US federal government since Executive Order 14028, and the European Cyber Resilience Act requires manufacturers of digital products to document their components.
SCRUM
Methodology= agility in sprints
Scrum is an agile framework that organizes product development into short, regular cycles called sprints (often two weeks), at the end of which the team delivers a usable increment. It was formalized by Ken Schwaber and Jeff Sutherland in the mid-1990s, drawing on a paper by Takeuchi and Nonaka on product development.
Scrum relies on three roles:
- the Product Owner, who prioritizes the backlog;
- the Scrum Master, who keeps the method alive and removes obstacles;
- the developers, who build the product.The pace is set by fixed events: sprint planning, daily stand-up (daily), sprint review and retrospective. Everything rests on three principles: transparency, inspection and adaptation. Scrum remains the most widespread agile approach, often blended with Kanban.
SCRUM Master
Role= the agile team coach
The Scrum Master is the person in charge of making Scrum work within a team: they help the team apply the framework, run the events and remove obstacles that slow delivery. They are neither a project manager nor a line manager: their authority comes from method, not title.
In practice, they facilitate sprint planning, the daily, the review and the retrospective, shield the team from outside interruptions, help the Product Owner keep a clear backlog and support the organization in its shift to agile.
The role is found mostly in large companies and organizations structuring several agile teams. In smaller companies, it is often taken on by an Engineering Manager or a Tech Lead.
Server Components
TermTech= React components that stay on the server
React Server Components (RSC) are React components that run only on the server and whose code is never sent to the browser. They have been a stable part of React since version 19.
A Server Component can read directly from a database, a file or a secret, then sends the browser a serialized description of what to display. Interactive parts (a button, a form) remain client components, explicitly declared with the "use client" directive.
The main benefit: less JavaScript to download and run on the user's side, so faster pages. Next.js, with its App Router, is the most widespread implementation. The model changes how a React application is designed: teams have to decide, component by component, what lives on the server and what lives in the browser.
Serverless
TechTerm= code with no server to manage
Serverless is a cloud execution model in which code is deployed without provisioning or managing servers: the platform allocates capacity on demand, automatically adjusts the number of instances and bills for actual usage.
There are still servers, but they are invisible to the team. The model comes in several forms:
- functions triggered by an event (AWS Lambda, Cloudflare Workers, Vercel Functions);
- serverless containers (Google Cloud Run);
- databases and managed services that scale on their own with load (DynamoDB, Aurora Serverless, Neon).Serverless works very well for irregular or unpredictable workloads. Under a constant, high load, it can cost more than a traditional infrastructure, and it comes with its own constraints: startup delay (cold start), limited execution time, strong dependency on the provider.
Service Mesh
TechTerm= the smart network between microservices
A service mesh is an infrastructure layer that handles communication between microservices: routing, encrypted traffic (mTLS), automatic retries, timeouts, traffic splitting and metrics collection, without changing application code.
In its classic form, a small proxy (often Envoy) runs next to each service as a sidecar. Every call goes through it. A central control plane distributes rules to all the proxies. Newer approaches drop the sidecar and move this work to the node or Linux kernel level (Istio's Ambient mode, eBPF-based Cilium).
The best-known implementations are Istio, Linkerd and Consul. A service mesh is justified when the number of services makes security and network visibility unmanageable by hand. On a small architecture, it mostly adds complexity.
Shape Up
Methodology= six-week cycles on a fixed budget
Shape Up is a product development method published in 2019 by Ryan Singer at Basecamp, which organizes work into six-week cycles with fixed time and variable scope.
Before each cycle, experienced people "shape" the topics: they frame the problem, sketch a solution without going into detail and set an appetite, meaning the amount of time the team is willing to spend on it. These proposals (pitches) are decided on at a betting table. A small team then takes the topic on with full autonomy, without split-up tickets or fine-grained estimates.
Each cycle is followed by a two-week cool-down to fix bugs, reduce debt or explore. Shape Up is mainly suited to compact, senior product teams that want to move away from the rhythm of Scrum sprints.
Singleton
Term= the one-of-a-kind object
The Singleton is a design pattern that ensures a class has only one instance across the whole program and provides a global access point to it. It is one of the patterns described in the reference "Gang of Four" book (1994).
It is used for resources that must remain unique: application configuration, a shared database connection, an event log (logger). In practice, the class constructor is made inaccessible and a method always returns the same instance, created on the first call.
The Singleton is also one of the most criticized patterns: it creates hidden global state, makes testing harder and tightly couples code. Many teams prefer dependency injection, which achieves the same result more explicitly.
Smart Contract
TechTerm= a program that enforces an agreement
A smart contract is a program deployed on a blockchain that runs automatically when its conditions are met, with no intermediary, and whose outcome can be verified by every participant in the network.
It is not a contract in the legal sense: it is code that holds assets and applies rules (transferring a token, releasing a payment, recording a vote). Smart contracts are the basic building block of DeFi, NFTs and DAOs. They are written in dedicated languages: Solidity for Ethereum and compatible networks, Rust for Solana, Move for Aptos and Sui.
Once deployed, a smart contract is in principle immutable, and it often handles real money. A logic error can be exploited within minutes. Hence the importance of security audits, thorough testing and bug bounty programs before going to production.
Snowflake / BigQuery
TechTool= the data warehouses of the cloud
Snowflake and BigQuery are cloud data warehouses: managed services that store very large volumes of data and let you query them in SQL, with storage and compute kept separate and billed by usage.
Snowflake is an independent vendor whose platform runs on AWS, Azure and Google Cloud. It isolates workloads in virtual warehouses that are started, stopped and sized on demand. BigQuery is Google Cloud's service: fully serverless, it is billed either by the amount of data read per query or by reserved capacity.
Both serve as the foundation of the modern data stack: ingestion tools load data into them, dbt transforms it, a BI tool presents it. They compete directly with lakehouse architectures such as Databricks.
SOA Architecture
TermMethodology= the system split into reusable services
Service-Oriented Architecture (SOA) is a way of organizing an information system into autonomous services, each delivering a business function (invoicing, checking a customer, computing a price) and communicating with the others through standardized interfaces over the network.
The idea is reuse: a single "customer" service serves the website, the call center and accounting, instead of each application coding its own version. In its classic form, from the 2000s, SOA often relies on a service bus (ESB) that routes and transforms messages, and on protocols such as SOAP.
It is found mostly in large information systems: banks, insurers, telecom operators, public administrations. Microservices are an evolution of it: smaller services, deployed independently, without a heavy central bus. Many modernization projects consist precisely of evolving an aging SOA.
SOC 2
TermConcept= the security passport of SaaS vendors
SOC 2 (System and Organization Controls 2) is an audit framework defined by the AICPA, the American association of certified public accountants, which attests that a service provider properly protects its customers' data.
The audit covers five criteria (Trust Services Criteria): security, always included, then availability, processing integrity, confidentiality and privacy, as chosen. An independent audit firm writes a report. Type I checks that controls are properly designed at a given date. Type II checks that they worked over a period of several months, and that is the one customers ask for.
It is not a certification but an attestation. SOC 2 is often required by large enterprises, especially in North America, before they entrust data to a SaaS vendor. In Europe, ISO 27001 plays a comparable role. Platforms such as Vanta or Drata automate part of the evidence collection.
Software Craftsmanship
MethodologyTerm= code built by the book
Software craftsmanship is a movement that treats software development as a craft: software must not only work, it must be well designed, readable and easy to evolve. It was formalized in a manifesto published in 2009, which extends the Agile Manifesto.
In practice, it relies on demanding technical practices:
- automated tests and TDD;
- code review and pair programming;
- continuous refactoring and design principles (SOLID, clean code);
- knowledge sharing and mentoring between developers.Agile organizes how a team delivers, craftsmanship deals with how code is written. The two complement each other: without technical quality, an agile team ends up slowed down by its debt.
Solid
FrameworkTech= fine-grained reactivity without a virtual DOM
Solid (or SolidJS) is an open source JavaScript library for building user interfaces, created by Ryan Carniato, which borrows React's syntax (JSX, components) but works without a virtual DOM, thanks to fine-grained reactivity based on signals.
In React, a state change re-runs the component's render. In Solid, each component runs only once: only the parts of the page that depend on a changed signal are updated. The result: very good performance and very little code sent to the browser.
SolidStart plays for Solid the role Next.js plays for React. The community remains much smaller than React's, but Solid helped popularize signals, an approach since adopted in various forms by Angular, Svelte, Preact and Vue.
Sovereign cloud / SecNumCloud
TermConcept= cloud shielded from foreign laws
Sovereign cloud refers to a cloud offering whose data and operations stay under European jurisdiction, shielded from extraterritorial laws. SecNumCloud is the qualification issued by ANSSI, the French cybersecurity agency, which sets the strictest requirements for it in France.
The SecNumCloud framework imposes technical and organizational security rules, and its recent versions add criteria for immunity from non-European law: data location, ownership control, staff authorized to operate the service. Obtaining the qualification requires a long and demanding audit.
In France, using a qualified offering is expected for certain sensitive data of public administration and operators of vital importance. The topic also concerns healthcare, defense and industry. Hybrid offerings combine hyperscaler technologies with a European operator to target this qualification.
Spotify Model
Methodology= squads, tribes, chapters and guilds
The Spotify Model is a way of organizing product and tech teams, described in 2012 by Henrik Kniberg and Anders Ivarsson based on their experience at Spotify, built on autonomous teams grouped by domain and connected by communities of practice.
It is structured around four levels:
- squads, small cross-functional teams responsible for part of the product;
- tribes, which group several squads working on the same domain;
- chapters, which bring together people with the same skill across squads (for example all the back-end developers in a tribe);
- guilds, communities of interest open to everyone.The model has been widely copied, often to the letter. Yet its authors presented it as a snapshot at a given moment, not a framework to replicate, and Spotify itself has changed its organization since. It remains a useful source of inspiration, not a ready-made method.
SQLite
TechTool= the database in a single file
SQLite is a relational database engine written in C, distributed as a library embedded directly into an application, which stores the entire database in a single file and is queried with SQL. It needs no server to install or manage.
This lightness makes it one of the most widely deployed databases in the world: it is embedded in smartphones (iOS, Android), web browsers, many desktop applications and connected devices. Its code is in the public domain.
SQLite is ideal for local storage, mobile apps, testing and prototypes. It is less suited to many users writing at the same time, where a server such as PostgreSQL or MySQL is usually chosen. Some newer projects also use it server-side for lightweight applications.
SRE
Role= the production reliability engineer
An SRE (Site Reliability Engineer) is an engineer responsible for the reliability of systems in production, who treats operations as a software problem: availability, performance, capacity, incident management and automation of repetitive tasks.
The practice was born at Google in the early 2000s and popularized by the book Site Reliability Engineering (2016). It spread a vocabulary that has become common:
- SLI and SLO: the measured reliability indicator and the target to reach;
- error budget: the tolerated margin of unavailability, which arbitrates between new features and stabilization;
- toil: repetitive manual work, to be reduced;
- blameless post-mortems after each incident.SRE shares a lot with DevOps, but with a more precise goal: reliability that is quantified, measured and negotiated with product. At Google, the team must spend at least half its time on engineering work rather than pure operations.
Staff / Principal Engineer
Role= the technical expert beyond senior
Staff Engineer and Principal Engineer refer to engineering levels beyond Senior, on an expert track (individual contributor track) that lets engineers progress without becoming managers.
A Staff Engineer has an impact that extends beyond their team: they drive structural architecture decisions, unblock cross-team projects, set technical standards and help others grow. A Principal Engineer exerts the same influence across an entire organization. The content varies widely between companies: some stay very close to the code, others act as architects or as the CTO's right hand.
These titles come from large US tech companies, where they match precise levels in a career ladder. In France, they are mostly found in scale-ups. In a smaller company, the same role may be called Tech Lead or Architect.
Sublime Text
Tool= a fast code editor
Sublime Text is a proprietary code editor available on Windows, macOS and Linux, known for its speed, lightness and clean interface. It was created by Jon Skinner and first released in 2008.
It popularized several features that have become standard in modern editors: multi-cursor editing, the command palette and instant navigation across a project's files. It supports syntax highlighting for a large number of languages and can be extended with plugins written in Python.
The software is paid, but can be evaluated for free. Many developers have since moved to Visual Studio Code or more complete IDEs, but Sublime Text keeps a following for its speed on large files.
Supply Chain Attack
TermConcept= hack the supplier to reach the customer
A supply chain attack is an attack that compromises an upstream supplier, tool or component in order to reach, through it, every organization that uses it.
The entry point can be a booby-trapped software update, a tampered open source library, a build tool or a managed service provider. A few landmark cases:
- SolarWinds (2020): a compromised update distributed to thousands of customers, including government agencies;
- 3CX (2023): a booby-trapped telephony application, itself compromised through another third-party program;
- xz-utils (2024): a backdoor introduced after years of contributions by a malicious maintainer, discovered just before it spread widely.Countermeasures combine dependency inventories (SBOM), signing and traceability of artifacts (Sigstore, SLSA), isolated builds and fewer third-party dependencies.
Svelte
FrameworkTech= the framework that disappears at compile time
Svelte is an open source framework for building web interfaces, created by Rich Harris in 2016, which works as a compiler: it turns components into optimized JavaScript at build time, instead of shipping a framework engine to the browser.
Components are written in files close to HTML, with simple logic blocks and CSS scoped to each component. The code sent to the browser is light and fast. Svelte 5 introduced runes, a reactivity model based on signals, which makes behavior more explicit in complex applications.
SvelteKit is the ecosystem's application framework (routing, server rendering, deployment), the equivalent of Next.js for React. Svelte regularly ranks among developers' favorite tools, but its adoption in companies remains well below that of React or Vue.
Symfony
FrameworkTech= the frame of PHP projects
Symfony is an open source PHP framework, created by the French company SensioLabs and its founder Fabien Potencier, released in 2005, which provides a structure and reusable components for building web applications and APIs. It follows an MVC-style architecture.
It is also a set of standalone components (routing, forms, security, console, dependency injection) reused by other projects: Laravel and Drupal partly rely on Symfony components. The Doctrine ORM is very often paired with it for database access.
Symfony is particularly present in France and Europe, in business applications, e-commerce platforms and complex back-ends. Its release cycle is regular and documented, with long-term support versions, which reassures projects that need to last.
Synchronous / Asynchronous
TermConcept= waiting for the answer or not
An exchange is synchronous when the sender waits for the response before continuing, and asynchronous when it carries on and handles the response later, whenever it arrives. The distinction applies both to code and to ways of working.
In programming: a synchronous call blocks the program until the result comes back. An asynchronous call starts the task (network request, file read, API call) and lets the program continue. This is essential to keep an interface smooth or let a server handle many requests in parallel. JavaScript and Node.js rely heavily on this model, and message queues such as Kafka make it possible to decouple services.
In organizations: a meeting or a call is synchronous, an email or a ticket is asynchronous. A messaging tool such as Slack is in principle asynchronous, even if it feels real-time. Remote teams rely heavily on asynchronous writing.
Tailwind CSS
FrameworkTech= styling written right in the HTML
Tailwind CSS is an open source utility-first CSS framework: instead of writing separate stylesheets, you build an element's appearance directly in the HTML with small utility classes (margin, color, font size, layout).
A build tool scans the code and generates only the classes actually used, which yields a lightweight final CSS file. Colors, spacing and sizes are defined in a single configuration that serves as the basis of the design system. Version 4, released in 2025, rebuilt the engine for speed and moved that configuration directly into CSS.
Its supporters like no longer having to name classes or worry about orphaned styles. Its critics point to cluttered HTML. Tailwind is very widespread in React and Next.js projects, often paired with ready-made component libraries.
TDD
Methodology= the test before the code
TDD (Test-Driven Development) is a development method where you first write an automated test describing the expected behavior, then the minimal code to make it pass, before improving that code. It was popularized by Kent Beck, in the wake of Extreme Programming.
The cycle has three steps, often called "red, green, refactor":
- write a failing test (red);
- write just enough code to make it pass (green);
- clean up the code without breaking the tests (refactor).TDD forces the need to be clarified before coding, limits unnecessary code and produces a test suite that secures future changes. Its cost: a slower start and a discipline that not every team sustains over time.
Terraform
ToolTech= infrastructure described in files
Terraform is an infrastructure as code tool created by HashiCorp in 2014, which lets you describe an infrastructure (servers, networks, databases, accounts, DNS) in text files, then create and modify it automatically with one or more cloud providers.
You write the desired state in a declarative language, HCL. The plan command computes the gap with what exists and shows what will change; apply applies the changes. Terraform keeps track of what it manages in a state file, which has to be stored and protected carefully. Hundreds of providers cover AWS, Azure, Google Cloud, OVHcloud, Scaleway, but also services such as Cloudflare or Datadog.
In 2023, HashiCorp moved Terraform to a non open source license. Part of the community then created OpenTofu, an open source fork hosted by the Linux Foundation and compatible with existing configurations. HashiCorp has since been acquired by IBM.
Token (AI)
TermConcept= the piece of a word
A token is the unit of text that a language model (LLM) works with: a fragment that can be a short word, part of a word, a punctuation mark or a character, produced by splitting the text with a tokenizer before passing it to the model.
The model does not read letters or words, but a sequence of tokens, and it generates its answer token by token. The token is the unit of measure everywhere: the size of the context window, generation speed and API billing, which usually distinguishes input tokens from output tokens.
Each model family has its own tokenizer. The same sentence therefore does not yield the same number of tokens from one model to another. Because tokenizers are often optimized for English, a text in French or a less represented language generally uses more tokens, so it costs more and takes up more context.
Training (AI)
TermConcept= the model learning phase
Training is the phase during which an AI model learns from data, gradually adjusting its internal parameters to reduce its errors on a given task.
The principle: the model is shown a large number of examples, the gap between its answer and the expected answer is measured, then its parameters are corrected in the right direction. The operation is repeated millions of times. For a large language model, a distinction is made between pre-training (learning language on a massive corpus) and later stages such as fine-tuning or RLHF, which steer it toward a use case.
Training is the opposite of inference, which means using the model once it has learned. Pre-training a foundation model requires GPU clusters and highly specialized teams; most companies simply adapt an existing model.
Transformer
TechTerm= the architecture behind LLMs
The transformer is a neural network architecture introduced by Google researchers in 2017 in the paper Attention Is All You Need, which underpins nearly all large language models and many image, audio and multimodal models.
Its central idea is the attention mechanism: to process each element in a sequence, the model weighs the importance of all the other elements. Unlike the recurrent networks that came before it, it processes the sequence in one block rather than word by word, which makes it possible to train very large models in parallel on thousands of GPUs.
GPT, BERT, Llama, Claude, Gemini and Mistral all rely on variants of the transformer. Its weak point: the cost of attention grows steeply with text length. Alternative or hybrid architectures (state space models such as Mamba) aim to reduce it.
tRPC
FrameworkTech= end-to-end typed APIs
tRPC is an open source TypeScript library for building an API between a server and a front-end application with types shared end to end, without writing a schema or generating code.
On the server side, you define procedures (queries, mutations, subscriptions). On the client side, you call them like local functions: the editor offers autocompletion, and any mismatch between the two sides is caught before the code even runs. Incoming data is validated with a library such as Zod.
tRPC is particularly well suited to monorepos where front-end and back-end are written in TypeScript by the same team, for example with Next.js. It is less relevant for a public API consumed by clients written in other languages: REST with OpenAPI or GraphQL are then better suited.
Trunk-Based Development
Methodology= everyone on the same branch
Trunk-based development is a way of organizing code in which all developers merge their changes into a single main branch (the trunk, often named main) at least once a day, instead of working for weeks on separate branches.
Branches, when there are any, live for a few hours. This requires three conditions:
- reliable continuous integration, with automated tests on every change;
- feature flags to ship unfinished code without turning it on for users;
- fast code reviews, so as not to block others.The benefit: avoiding painful merges of branches that have drifted too far apart, and being able to ship to production at any time. DORA research on software team performance associates this practice with the teams that ship most often and most smoothly.
Turborepo
ToolTech= the conductor of JavaScript monorepos
Turborepo is an open source task orchestration tool for JavaScript and TypeScript monorepos, acquired by Vercel in 2021, which speeds up builds, tests and checks by only re-running what has changed.
It reads the dependencies between the packages in the repository, runs tasks in the right order and in parallel where possible, and caches their results. If a package has not changed, its build is not re-run: the result is fetched from the cache, locally or from a cache shared between developers and CI. On a large monorepo, continuous integration time drops significantly.
Its main competitor is Nx, more complete but more opinionated. Turborepo bets on a minimal configuration that fits onto an existing repository. For very large, multi-language monorepos, tools such as Bazel take over.
TypeScript
LanguageTech= JavaScript with types
TypeScript is an open source programming language created by Microsoft in 2012, which adds a static type system to JavaScript. TypeScript code is converted into standard JavaScript that runs in any browser or in an environment such as Node.js.
Types describe the shape of data: what a function expects, what it returns, which fields an object contains. The compiler flags inconsistencies before the code runs, the editor offers precise autocompletion, and changes across a large codebase become much safer. Typing is gradual: an existing JavaScript project can be migrated file by file.
TypeScript has become the default choice for professional JavaScript projects, on the front end (React, Angular, Vue) as well as on the back end (Node.js, Deno, Bun).
UX/UI Designer
Role= the architect of user experience
The UX/UI Designer is the designer who shapes both the experience of a digital product (UX, user experience: journeys, logic, ease of use) and its visual interface (UI, user interface: screens, components, typography, colors). The title combines two specialties that are often separate in larger teams.
On the UX side, they interview users, map journeys, design wireframes and test prototypes. On the UI side, they produce the final screens and often maintain a design system shared with front-end developers, most of the time in Figma.
The combined profile is common in startups and small companies, where a single designer covers everything. As teams grow, they often split the roles (UX Designer, UI Designer, UX Researcher) or group them under the Product Designer title.
Vector Database
TechTool= the database that searches by meaning
A vector database is a database specialized in storing vectors, called embeddings, and in quickly finding the vectors closest to a query, which makes it possible to retrieve content by closeness of meaning rather than by keywords.
Each document, paragraph or image is converted into a vector by an embedding model. For a given question, the database returns the items whose vectors are closest. Over millions of vectors, an exhaustive comparison would be too slow: it uses approximate search indexes (HNSW, IVF, DiskANN) that trade a little precision for a lot of speed.
It is the typical storage layer of RAG architectures, recommendation engines and semantic search. There are dedicated databases (Pinecone as a managed service, Qdrant, Weaviate, Milvus or Chroma as open source) and extensions of existing databases, such as pgvector for PostgreSQL or the vector search in Elasticsearch and OpenSearch.
Vite
ToolTech= the ultra-fast front-end build tool
Vite is an open source front-end build tool created by Evan You, the author of Vue.js, in 2020, which provides both a near-instant development server and optimized compilation of code for production.
In development, Vite relies on the browser's native JavaScript modules: it serves files on demand instead of bundling everything before each launch. The server starts in moments and each change shows up on screen almost immediately. For production, it produces optimized, bundled files. Rolldown, a bundler written in Rust, is designed to unify and replace Vite's historical building blocks.
Vite works with React, Vue, Svelte, Solid and most front-end frameworks. It also serves as the foundation for frameworks such as Nuxt, SvelteKit and Astro, and for the Vitest testing tool.
vLLM / TGI / SGLang
ToolTech= the engines that run LLMs
vLLM, TGI (Text Generation Inference) and SGLang are open-source inference engines used to run a large language model quickly and efficiently on GPUs, in production.
Running an LLM naively wastes a lot of memory and compute time. These engines apply optimizations such as continuous batching of requests, fine-grained GPU memory management (vLLM popularized the PagedAttention technique), support for quantized models or splitting a model across several cards. They usually expose an API compatible with OpenAI's.
vLLM came out of research work at UC Berkeley, TGI is developed by Hugging Face, and SGLang also comes from academic research. They are core building blocks of inference platforms built on open-weight models.
VP of Engineering
Role= the head of the tech organization
The VP of Engineering is the executive in charge of organizing engineering teams: they make sure the teams deliver well, fast and sustainably, by driving hiring, team structure, processes and the management of managers. They usually report to the CTO or the CEO.
While the CTO owns the technology vision and major architecture choices, the VP of Engineering handles execution: delivery capacity, team organization, career paths, quality, budget. In some companies, one person holds both roles.
The role appears when the tech team grows beyond what a CTO can lead alone, typically once there are several teams and several Engineering Managers.
VueJS
FrameworkTech= the progressive front-end framework
Vue.js (or Vue) is an open source JavaScript framework created by Evan You and released in 2014, used to build user interfaces and single-page web applications (SPAs). It is designed to be adopted progressively: you can add it to a single part of an existing page or build an entire application with it.
Like React, it works with components that react to data changes. It stands out with component files that bundle the HTML template, logic and styles, and with a learning curve known to be gentle. Its official ecosystem includes Vue Router (navigation), Pinia (state management) and Vite (a build tool created by the same author). Nuxt is the application framework built on Vue.
Vue has a strong community in Europe and Asia, and is widely used by agencies, SaaS vendors and projects modernizing an existing application.
Wallet
TermTech= the key ring of the blockchain
A wallet is software or a device that stores a user's private cryptographic keys and lets them interact with a blockchain: receive and send crypto assets, sign transactions, connect to decentralized applications.
The wallet does not hold the assets themselves, which remain recorded on the blockchain: it holds the keys that prove ownership. There are:
- hot wallets, software connected to the Internet (MetaMask, Rabby, Phantom), convenient for daily use;
- cold wallets, physical devices that keep keys offline (Ledger, Trezor), safer for large amounts;
- smart wallets, programmable wallets (account abstraction) that allow account recovery, multiple signatures or fees paid by a third party.Losing one's recovery phrase (seed phrase) means permanently losing access to one's assets.
Web Architecture
TermConcept= the blueprint of a web application
Web architecture is the organization of a web application into building blocks (interface, server, database, external services) and the way those blocks communicate. It is decided early, because it shapes performance, security, hosting costs and how easily the product can evolve.
The basic pattern stays the same: a front end in the browser or on mobile, calling a back end through APIs, which reads from and writes to one or more databases. Around it come a cache, a message queue, a CDN, authentication or payment services.
The structural choices concern how the system is split (monolith or microservices), where rendering happens (server side or client side), hosting (servers, containers, serverless) and how load growth is handled. There is no good architecture in absolute terms: a startup looking for its market does not have the same needs as a platform with millions of users.
WebAssembly (Wasm)
TechTerm= native code in the browser
WebAssembly (Wasm) is a portable binary code format, standardized by the W3C, which lets programs written in languages such as Rust, C, C++ or Go run in a browser at near-native speed.
You do not write Wasm directly: you compile to it. The code runs in an isolated environment alongside JavaScript, with which it exchanges data. This is what made it possible to bring heavy software, once limited to desktop machines, into the browser: the web versions of Figma, Photoshop and AutoCAD are examples.
Wasm is also used outside the browser. Its fast startup and isolation make it a popular format for running code on edge platforms (Fastly Compute, Cloudflare Workers) or for plugin systems (Envoy, Shopify Functions). The WASI standard gives it controlled access to files and the network outside the browser.
Zero Trust
MethodologyConcept= never trust, always verify
Zero Trust is a security model in which no user, device or traffic flow is considered trustworthy by default, even when it sits inside the company network: every access to a resource must be verified.
It is the opposite of the perimeter model, in which a firewall and a VPN protect an internal network where everyone trusts each other. With remote work and the cloud, that perimeter no longer really exists. Zero Trust rests on three principles:
- explicitly verifying identity, with strong authentication (MFA);
- taking context into account: up-to-date device, location, unusual behavior;
- granting only what is strictly necessary (least privilege).The model was formalized notably by NIST and popularized by Google's BeyondCorp project. It is implemented in stages, often with network access solutions known as ZTNA (Zscaler, Cloudflare, Tailscale) that gradually replace VPNs.