AI pilot · Claude · Brussels & Luxembourg

I direct AI.
And I teach you to direct it.

AI writes code, rephrases, translates, moderates. But it decides nothing. A web developer for 25 years in Brussels and Luxembourg, a lecturer at Haute École Francisco Ferrer, I build AI systems (Claude) that actually run in production, and I train those who want to pilot AI instead of enduring it.

Work together Get AI training

this very site is built by directing Claude ↓

AI is the engine.
Without a pilot, the car ends up in the ditch.

AI produces code faster than any developer. It drafts terms and conditions, replies to clients in Dutch, turns a mockup into a web page. What it does not do: know when it is wrong.

For that, you need a pilot. Someone who understands architecture, security, databases, deployment. Not to do everything by hand: to validate, correct, decide. That is exactly what 25 years in the field taught me, and it is what separates an impressive demo from a product that holds up in production.

Part 1 · Practitioner

I build with AI

Not conference prototypes. Systems with real users, real outages and real legal constraints.

01
Sillage

An AI that moderates without censoring

An ephemeral journal of messages left in Brussels public transport: a three-stage moderation pipeline (instant local filter, OpenAI moderation API, then Claude Haiku for emotional nuance). The prompt separates real hostility, which is filtered, from sadness, social criticism or dark humour, which pass. If one AI goes down, the message still gets through: the system is designed for failure, not for the ideal case.

02
PlateWave · No. 1 App Store Luxembourg

The AI that changes the tone, not the substance

My app PlateWave lets drivers message each other via their licence plate. Between sending and receiving, an AI rephrases: raw on the way out, courteous on the way in. The factual content stays intact, only the tone changes. That is the idea that unlocked a project stuck in a drawer for 15 years: without automatic moderation, it would have turned into insults within 48 hours.

03
Architecture

The right model in the right place

I do not believe in the single model that does everything. Claude Haiku for fast, cheap nuance, GPT-4o-mini for text transformation, the Moderation API for raw filtering. Choosing is already architecture.

04
Production

The whole chain, not one link

Prompt engineering, API integration, orchestration, monitoring, GDPR compliance, going live. I do not ship a script that calls ChatGPT: I ship a complete, documented system that survives its own launch.

The simplest proof: you are on it.

This site is built by directing AI. The sections you are browsing were specified, generated, checked screen by screen, then corrected, in short loops where the AI produces and I decide. Even the inventory of my AI skills was not written from memory: it was extracted from a knowledge graph built on twelve of my projects, so I only claim what really exists in my code.

That is my method, and that is what I sell: not magic, a discipline. A clear context, structured knowledge, validation loops, a project memory.

AI does the volume. I make the difference.

Part 2 · Trainer

I teach how to direct AI

At Haute École Francisco Ferrer, I redesigned the third-year curriculum around a simple idea: train pilots, not passengers.

UE 501 6 ECTS 12 weeks

UE 501: the engine room

I created and teach the unit "AI-pilot development" (6 ECTS). Twelve weeks of project work over a full cycle: design, specify, generate, validate, deploy. Students do not learn to type code faster. They learn to make quality code get produced, and to recognise when it is not.

The 4 pillars of the AI pilot

  1. 01

    Context

    Structure what the agent needs to know (CLAUDE.md, token budget).

  2. 02

    Knowledge

    Organise the project for AI (structured repo, knowledge graph).

  3. 03

    The work loops

    Generate, validate, correct, repeat.

  4. 04

    Memory

    Make several AI sessions collaborate without overwriting each other.

Figma + Claude + MCP: from design to code

A full specialisation in six sessions: AI brainstorming, wireframing, Auto Layout in Figma, AI as a CSS assistant, animation and prototyping, then moving from design to code. Thanks to the MCP protocol, the agent reads the Figma mockups directly and generates faithful code: frame, generation, side-by-side comparison, correction. I compare three Figma MCP servers there, and I also teach the WordPress MCP, with which the agent creates pages and content directly in the CMS.

My own tools

I write my own Claude Code skills: WordPress block development, performance auditing, static analysis. When a tool does not exist, I build it. And my students learn with it.

What I tell my students

The question is no longer "can I type code?" but "do I understand what the AI is building for me?".

Generated code is often functional and often insecure. The skill that matters: spotting and fixing what the AI lets slip through.

Two ways to work with me

You have a project

A product where AI belongs? A website, an app, an internal tool? I design it, I build it, I put it into production.

Let's talk about your project

You want to get trained

Are you a developer, a teacher, or the head of a technical team? I train people to pilot AI: Claude Code, MCP, agentic workflows. On site, remotely or one-to-one.

Book a training