Lawyer by training. AI systems, law, cognitive sciences. Open publications and tools.
In the Legal Prompting Manual, I wrote that the point was not to treat artificial intelligence as a threat, but to grasp one thing: your choices will shape reality. Either AI will serve you, or you will end up serving AI.
I defend the vision Garry Kasparov formulated after his 1997 loss to Deep Blue: the centaur. A tandem in which human and machine play together, each amplifying the other's capacities. AI serving the human, alongside the human, and never without.
What worries me is not what AI can do. It's what it is already doing, silently, to those who use it.
Law studies in Brest, Master's in Justice, Trial and Procedures in Rennes, graduate diploma in Judicial Careers at Paris 1 Panthéon‑Sorbonne. CRFPA (bar school entrance exam) in criminal law, admissibility to the French Magistracy Contest. Twenty months at the Prosecutor's Office of Rennes, drafting indictments, appeal briefs and legal summaries alongside magistrates.
That's where I started looking into artificial intelligence. Not out of fascination with the technology, but because I saw tools arriving in professional environments without anyone really understanding how they worked, or where they broke.
Since 2023, I design AI systems: prompt architectures, anti‑hallucination protocols, behavioural model evaluation. I'm not a software engineer. What I bring is the rigour of legal reasoning to a field that often lacks it.
Prolonged interaction with a language model produces a profiling effect that goes beyond the risks usually documented. Working paper formalising three concepts: user imprint, qualified information (extending Zuboff's framework to the conversational context), and operational response.
Read summary →A 46‑page manual for magistrates, lawyers and judicial auditors. How LLMs actually work, cognitive biases, professional secrecy under AI, a structured prompting method. Nearly half the audience came from the Ministry of Justice, the National School for the Judiciary, or the Paris Bar.
Read summary →Open‑source multi‑model architecture for legal analysis. Calibration externalised on Haiku, tagged analysis on Sonnet or Opus, active verification via the CoVe protocol. Five epistemic tagging levels, seven MCP tools connected to Judilibre, a 17‑domain transversality matrix.
Read summary →My advisory work funds the research. The medium‑term objective is a doctorate at the intersection of AI, law and cognitive sciences. I'm not there yet. But every project, every publication, every question builds the path.
The work listed here is public and discussable. So are the open threads.
Book a call