AI · Data · Ethics · CreationContact

Selected projects · method in action

A question. Then the missing instrument.

These projects follow the same approach: identify a blind spot, design a useful response, test it and bring it into production.

First case — ScoreMyDataCenter ↓

Flagship project · 01

ScoreMyDataCenter, the missing rating instrument.

Can the social acceptability of data centers be measured?

Data centers are multiplying and their critical risk is also territorial: water, energy, land, local impact and transparency. Major projects collide with this risk, without a public instrument to make it readable. I built an acceptability grade from for every site, grounded in a public method and open data.

Explore ScoreMyDataCenter ↗
Eurofiber · Rotterdam, Netherlands — A
Eurofiber · RotterdamNetherlands
DataSix · Vienna, Austria — B
DataSix · ViennaAustria
AntwerpDC · Antwerp, Belgium — C
AntwerpDC · AntwerpBelgium
AtlasEdge · Brussels, Belgium — D
AtlasEdge · BrusselsBelgium
Campus IA · Fouju, France — E
Campus IA · FoujuFrance
Explore the interactive map of graded sites ↗
01

The observation

See the risk existing tools do not measure.

Standards describe the technical robustness of infrastructure. They do not answer the decisive question: can this project take root sustainably in its territory?

02

The angle

Borrow the form that makes the problem visible.

The spark came from Nutri-Score: a simple letter can reshape how a complex market is read. I applied that principle to something nobody knows how to compare this way: a data center in its territory.

03

Design decisions

Method before announcement.
  • A five-letter scale rather than a score out of 100: readability before false precision.
  • A grade always paired with its evidence level: we do not grade the unknown.
  • For a future project, a provisional grade that narrows as data arrives.
  • A public method, open data and every value linked to its source.
04

Execution

Move from framing note to a production system.

In one summer and solo, with an orchestrated system of AI agents: public-source collection, cleaning, audit, rating engine, project sheets, rankings, country indices, reports and API. More than 1,400 European sites are now covered.

CollectStructureAuditAnalysePublish
05

What it changes

A scattered reality becomes comparable and contestable.

A territory, operator, journalist or investor can now start from a legible grade, trace it back to the facts and debate the method. The instrument does not erase debate: it gives it common ground.

A judge of transparency cannot be a black box.

The proof, directly

ScoreMyDataCenter

Campus IA FoujuE for the site · C for the projectCampus IA de BosquelA provisional grade that narrowsMethodologyThe public, verifiable frameworkAPIReuse the data and scores

Other projects

Same approach, other fields.

02

AI security · AI RiskPath

AI RiskPath: know what still protects AI systems.

When AI can search, code, call tools and act in chains, the question is no longer only “how capable is it?”. The real question becomes: what still prevents it from doing harm? A provider closing an account, a model kept under lock, a human code review, a long sequence of steps, a trusted-user gate: these locks are often outside your product, and they can give way without warning. AI RiskPath dates these safety dependencies to show what still holds, what has already been lifted and what nobody has really tested.

Why it matters
  1. For everyone: AI safety is not a magic switch, but a set of concrete locks that can weaken over time.
  2. For organisations: do not build critical uses on external defences without knowing who holds them and since when.
  3. For decision-makers: separate what has been observed from what is still assumed or untested.
  4. For technical teams: turn vague worries into auditable questions, tests and guardrails.
03

Industrialise knowledge

From technical documentation to a repair decision.

Two products apply the same method to fragmented technical knowledge. For dishwashers, EDGARD — Entretien Domestique : Guide d’Aide aux Réparations Durables — gives the promise a face: repair before replacing. Promote durability.

The shared method: a multi-agent system
  1. Collect official manufacturer documentation and documented technical cases.
  2. Use AI to extract and normalise codes, causes, parts and values while preserving source traceability.
  3. Turn diagnosis into clear decisions: check, repair it yourself, order the right part or call a professional.
  4. Separate safe actions, professional interventions and warranty cases.
04

Make a concept tangible

Git Demo Interactive

A simulation for understanding how GitHub works visually. Branches, commits and merges — with no installation.

05

Turn an intuition into research

How Words Bend AI Models

Do you know the staircase test?

  1. Give an AI a photograph.
  2. Ask it exactly the same underlying question while changing the tone: once with concern, once with enthusiasm. The facts do not change, but the answer flips.
  3. Here is an example using a staircase.
Alarmist tone
Question

‘They creak terribly. They are dangerous.’

AI answer

Yes, this staircase may present a real danger. The creaking could signal looseness or wear: it would be best to have it inspected and repaired quickly.

Enthusiastic tone
Question

‘I love them. Such character! The creak feels authentic.’

AI answer

Yes, the creaking can reinforce the authentic charm of a wooden staircase. If it remains stable, with no looseness or cracks, you can absolutely keep it as it is.

Same staircase. Same AI. Two opposite assessments.

AI does not give you the right answer, but the one that matches your mood!

When the advice concerns money, health or a life decision, this invisible — and still poorly documented — bias becomes a risk.

This is the focus of the tonal research I have been conducting since 2024.

Launch the interactive viewer ↗

Original research · reproducible protocol · scientific publication

06

Document intelligence

Collège de France ‘AI 2025’ conference: using AI to work through a complex corpus.

Drawing on the talks from this multidisciplinary symposium and their supporting documents, a synthesis produced through advanced document prompting and context engineering — far beyond a simple ‘summarise this’.

Summarise without distorting; preserve disagreements, nuance and texture: reducing is not rewriting.

Dozens of hours of talks. One editorial constraint: a two-page A4 brief for each speaker.