PARIS · INDEPENDENT DATA & AI CONSULTANT
I design and productionise GenAI agents that actually work with your teams: legal, sales, healthcare. On Google Cloud, with measurable guardrails.
Available now
Trusted by
OFFERS
A short format to decide, a medium one to deliver, a long one to make it last. Day rate shown, each offer priced on request.
5 days
Know where you stand in 5 days, and what to do next.
For a team that already has a POC, a pilot or a clear idea, and wants to decide based on facts.
4 to 6 weeks
An agent in production in 4 to 6 weeks.
For a business unit or a product team with a well-scoped use case and users ready to adopt it.
Long-term engagement
Productionise, govern, upskill your teams.
For a CTO or Head of Data who wants to go from one agent to several, without multiplying the risk.
Day rate €600 excl. VAT / day
Each offer is quoted on request, after an initial 30-minute conversation.
FROM POC TO REAL USE
An agent that impresses in a meeting is useless if nobody is using it a month later. I build the other way round: usage first, then the model.
We pick a use case your teams already handle by hand, and we pin down what ‘success’ means.
A prototype on your real data, judged by an evaluation set, not by a gut feeling.
The agent goes live on Google Cloud, held to the same standards as a critical service.
The agent learns from your teams’ feedback and stays under their control.
$ agent run --task "contract review" --ref internal-playbook [plan] 3 steps: extract clauses → compare against playbook → propose revisions [tool] docx.extract_clauses(document) → 42 clauses, 0 parsing errors [tool] playbook.search(clause="limitation of liability") → 2 internal rules [result] gap detected: no cap, rule R-12 requires an explicit cap [tool] memory.get(lawyer, org) → 1 preference, 3 senior-validated patterns [check] verbatim quote: cited passage ⊂ source document ✓ [guard] clause 17: source not found → revision blocked, flagged to the lawyer [judge] A/B pairwise vs previous version → preferred in 7/9 cases [tool] docx.write_tracked_changes(revisions=11) → OK [trace] langfuse span 2f9a… · 6 LLM calls · 14.2 s · cost recorded [answer] 11 revisions proposed, 1 point for the lawyer to decide. Nothing invented.
PROJECTS
Nine projects, from medical imaging to GenAI agents. Each one with the problem, what I did, and what came out of it.
Contract review is slow and inconsistent: every lawyer applies the internal playbook their own way, with no audit trail and no memory.
Result Used daily by the legal team, with a public online demo.
Statements of Work take too long to produce and often break the funding rules: caps, fair market value, formats.
Result In production for managers and sales teams.
Preparing an outreach brief on a scientific key opinion leader takes hours, and an LLM on its own invents references.
Result 0 hallucinated citations across the briefs produced.
Emergency dispatchers have seconds to triage a call, with partial information and lives at stake.
Result Classification into 5 emergency levels and a continuous stroke / heart attack risk score.
On a large corpus mixing images and text, search returned results of poor relevance, with no reproducible way to evaluate them.
Result Measurable retrieval improvement on the evaluation protocol.
Patient data is siloed across practice management software, making it impossible to use as a whole.
Result POC validated and operational handover to the teams.
EXPERTISE
I learned rigour in a field where AI has no room for approximation: healthcare, FDA, imaging. Today I apply it to your teams’ GenAI agents.
Designing agents that abstain rather than invent.
Deploy, trace, secure and maintain on Google Cloud.
Vision, signal, text: models trained to run in real time.
Standards, regulatory submissions and clinical trials.
CAREER
Each step added a new requirement: standards, clinical evidence, data, real time, then taking agents to production.
HOW I WORK
A prototype only counts if it has named users and a production date.
I scope the use case before the model, and I measure adoption afterwards.
The agent that produces is never the one that verifies: a sceptic reviews, and programmatic verification has the final say.
This pattern is what delivered 0 hallucinated citations on the KOL multi-agent system.
Prompt, tool called, cost, latency, judge verdict: everything is in Langfuse and Cloud Trace.
When an agent gets it wrong, you know where, why, and how much it cost.
Compliance rules are coded deterministically, preferences come from senior staff, the agent applies them.
When a source is missing, it abstains and hands over to a human.
CONTACT
Thirty minutes is enough to know whether your use case holds up, and which offer to start with.
Available now
Book 30 minutes (opens in a new tab)