Speciality
AI does the work. Judgment stays human.
No "future of AI" slides here: a pipeline running in production, tokens generated then hand-checked, and working habits that go through the data before the interface.
01
AI as a product engine
Two production pipelines: ingestion (the city meshed in H3 hexagons, swept through Google Places) and the mention scanner, which has digested 40,000+ editorial mentions: GPT-5 classifies each mention, Claude Haiku generates tags and summaries. Above the database, two multi-agent systems: editorial writing from our data, and discovery of new places. And the product itself is built 100% with Claude Code: I didn't type a single line of code: I decide, test and correct. A deliberate precision/recall trade-off: places under 4.3 on Google are not scanned.
Choosing what you leave out is how you protect the quality of what you show.
Read the case study: Gatto02
A multi-agent system for conformant DSFR
Every French State service comes with the same imposed base: strict DSFR, RGAA. Reproducing it by hand on every screen wastes time and invites conformance drift. I built a React boilerplate that embeds the DSFR guidelines, coupled with a generation flow driven by specialised agents (UX research, UI design, code review) intervening at every iteration. The designer describes a journey; the system produces conformant screens, ready to be refined. React rather than generating through the Figma MCP: on full journeys, generating straight into code shortens the loop between design intent and a screen you can manipulate. A 100% pure DSFR base, versioned on Git, made to be handed over. Used on a real PoC, then picked up by my colleagues: handover without documentation, live at the office (install, how it works, first screens). On a State pre-sales bid, the boilerplate generated dozens of conformant screens in two weeks, used as-is. A journey goes from several days to a few hours.
The real proof of a tool: it passes from hand to hand without me. Conformance rides on the DSFR base, not a last-minute check.
Read the case study: DSFR boilerplate, ANS missions03
AI as a systems tool
Generating dozens of DSFR-compatible JSON design-token variants (the combinatorial, repetitive part), then checking everything by hand: contrast (RGAA thresholds), brand fidelity, cross-component coherence, accessibility non-regression.
AI does the combinatorial; judgment stays human. Never the other way round.
Read the case study: ANS × DSFR Design System04
AI as a production studio
A promotional video for an internal project, CeSaR: the script written with AI, the screens calibrated to the script, the voice-over generated with ElevenLabs. Edited in DaVinci Resolve, learned for the occasion.
AI didn’t replace the editor: it taught me how to become one.
Watch the video on YouTube05
AI as a way of working
Querying the Gatto database conversationally through MCP to validate a hypothesis before drawing the UI. Generating social content storyboards from verified SQL queries, not hunches. Using image generation as exploration material, never as a deliverable. And this portfolio itself is built in collaboration with AI: reviewed, arbitrated and corrected by a human who contradicts it often.
Database-first: test an idea against real data before building the interface.
What experience taught me
- Two models, two jobs: pick a model for a task, not for its reputation.
- Distrust silent success: a pipeline that returns no error has not necessarily worked.
- Human validation is not a checkpoint. It is where judgment lives: contrast, brand, meaning.
- AI is never the message: whatever leaves the pipeline must stand on its own, without saying how it was made.