1
Business service
The default. Enterprise or API offerings whose terms exclude training on submitted data. Suitable for most projects, sensitive data aside.
I work with coding agents daily, on my own projects and on my clients'. This page says how, and above all what I commit to. It is public and dated so you can hold me to it.
AI does not change who signs the work. It changes how fast I explore an idea, not the standard for what I deliver.
My commitments
Your data trains no one
No client content goes into a free consumer service. I only use business offerings whose terms exclude training on submitted data, or models I host myself.
I match your requirements
If your hosting is sovereign, regulated or air-gapped, the tooling follows the same constraint, or I switch it off. The choice is yours.
I stay accountable for what I deliver
Everything I deliver, I have read, I understand and I can maintain without assistance. If I cannot explain a line, it does not ship.
What I never do
- Put a secret, a credential or an API key into a prompt.
- Send personal data to a third-party service without a legal basis and your written agreement.
- Ship code I have not read line by line.
- Claim a compliance I cannot demonstrate.
Three tiers, depending on your context
The tier is agreed during scoping, and revisited if the data perimeter changes.
2
Sovereign inference
Mistral models served from Scaleway or OVHcloud: French model, French host, French law. For public administrations, Albert, the French state's open-source model, hosted in France.
3
Self-hosted or offline
Open-weight models deployed on your own infrastructure, down to environments with no internet access.
Sovereignty has a price: SecNumCloud-qualified hosting commonly costs 30 to 50% more than a non-qualified hyperscaler. My job is to help you weigh that, not to tick a box.
The guardrails
An agent speeds up writing, not validation. On my projects: strict typing, mandatory lint and tests in CI, systematic review, and an AGENTS.md file that bounds what an agent may do in the repository. The code of this site is public, you can check.
Building for agents, not just with them
This site exposes a documented agent surface and a versioned access policy. Personal pages are explicitly out of scope: the rule is written once in the code, then applied to robots.txt, to the capability descriptor and to the index meant for models. Making a system legible to agents without making it say everything is an architecture exercise like any other.
/mcpMCP server, read-only tools/llms.txtIndex for models/llms-full.txtDetailed instructions/.well-known/agent.jsonCapability descriptor, boundaries included
Last updated: 25 September 2026