I’m an independent Data & AI consultant and founder based in Paris, France. I have spent 20 years working across IT, data and AI, with a practical interest in the point where architecture, delivery and operations meet.
My current focus is production GenAI: turning useful prototypes into systems that can be evaluated, observed, secured and maintained. I work on design choices around retrieval and agentic workflows, quality measurement, runtime visibility, access boundaries, and operating models. The aim is not novelty for its own sake; it is to make AI behavior understandable enough for teams to improve it and dependable enough for real use.
- Production GenAI architecture and the path from prototype to operation
- Repeatable evaluation at component and system level
- Observability for quality, cost, latency and failure analysis
- Security, data boundaries and controlled tool access
- open-fullscreenshot — a Chrome Manifest V3 extension for full-page, visible-area, element and region capture, using active-tab access without host permissions, network calls or telemetry.
- generative-engine-monitor — a tool for measuring brand visibility in AI answers across parametric and grounded responses, with explainable scoring, confidence intervals and replay.
- ftp-deploy-mcp — an MCP server that connects AI coding tools to FTP, FTPS and SFTP, with multi-server support, path isolation, read-only mode and dry runs.
- tierdecay — a Markdown-based model-routing approach that distills solved problem classes into reusable execution patterns for AI coding CLIs.
- harnessmeter — an offline profiler for agent harness instructions, skills, subagents and MCP tool schemas, designed to work without API keys or network access.
- claude-inc — a structured collection of 42 role-based skills across seven departments for Claude Code and other AI CLIs.
- Start with the problem, constraints and failure modes.
- Treat evaluation and observability as product requirements.
- Minimize privileges and data exposure; make consequential actions explicit.
- Prefer simple, reviewable components and measurable automation.
French is my native language; I work professionally in English.