I build applied-AI products around a simple question:

If AI existed from day one, would we still do it this way?

Most of the time, the answer is no — and that's what I like exploring: using AI to rethink a real need from the ground up, instead of bolting it onto the old way.

What I genuinely love is the loop: understand the real need → hack and build → break it → rebuild it better. Then again.

• scopegraph An ecosystem-aware scoping assistant: describe a project idea and it finds the existing context in a curated knowledge graph — systems, past decisions, inherited constraints — then shows the links, challenges the need, and generates a grounded scoping dossier. The LLM proposes; a deterministic runtime decides.

• FitMAS An autonomous coaching staff for real-life consistency — built around fatigue, missed sessions and setbacks, not a perfect plan. The LLM proposes, a deterministic verifier holds authority, and the coach reaches out on its own. Live as my daily dogfood.

• TaupIA An AI oral-exam examiner for French prépa maths. It challenges a student's reasoning and guides with questions like a real examiner — it never hands over the answer.

• use-case-assistant Turns a vague business AI idea into a structured, usable intake form — challenging unclear needs and asking for missing context before jumping to a solution.

Agentic context must be designed, not improvised.

I'm a heavy daily user of both Claude Code and Codex — I live in these tools, alternating between them constantly on real work. So this comes from practice, not theory.

I don't see agentic IDEs as better autocomplete — I see them as a new delivery layer that needs governance. And it isn't abstract for me: I work on AI adoption inside a regulated bank, where quality, security, auditability, and review are non-negotiable. The real question isn't whether agents can generate code; it's how they fit into a serious SDLC without becoming an uncontrolled source of truth.

It's not just about dropping an agent into the existing process — the process itself is up for redesign (if AI were there from day one, we wouldn't deliver software the same way). The deeper blocker, though, is making the enterprise legible to an agent: the business domain, the information system, the internal frameworks — mostly scattered, undocumented, legacy knowledge. Turning that into a designed source of truth an agent can boot from is, to me, where the real leverage is.

I don't claim answers; I'm immersed in the problem, daily — and some of my projects are where I test these ideas in the open, where state — not prose — is the proof.

Read the full note: agentic IDEs in a regulated SDLC →

I'm exploring applied-AI roles close to real problems, real users, and real products — especially around education, coaching, and enterprise AI workflows.