I build with each new AI capability, to make people more efficient. Sometimes that means putting it in the hands of people who never read a model card. Sometimes it means pushing it further than it goes alone.
Co-founder and Chief AI Officer at DeltaWave, Paris. I designed EVE, our AI co-scientist. A scientist asks whether a drug's binding is reproducible in docking; EVE finds the structures, runs the experiment, and returns an answer traced to its sources, in a secure environment the customer controls. First enterprise contract signed with a Nasdaq-listed company (deployment in progress). Second SPRIND Challenge award in 2026.
Before DeltaWave: data scientist at SESAMm and reciTAL. M2 in mathematics (M2MO, Université de Paris).
An Auditable Agent Platform for Automated Molecular Optimisation · co-author · Aug 2025
Six agents optimise molecules against the AKT1 protein with docking tools. Every tool call is logged, so each molecule keeps a traceable history.
- With Claude 3.7 Sonnet, the multi-agent setup reached the best mean docking score: −10.97 kcal/mol, against −10.65 for the LLM alone and −9.72 for a single agent (lower means stronger predicted binding).
- Single-agent runs gave more drug-like molecules: 100% pass Lipinski's rule of five, against 56% for multi-agent.
- This paper studies the first version of our harness.
Figure: Ünlü, Rohr, Celebi et al., arXiv 2508.03444, CC BY-NC-SA 4.0.
I build software with coding agents. I built these tools to make their work easier to inspect and review.
Earlier: Q-learning maze generator (2022), an RL agent that builds mazes. Write-up of the fine-tuning dataset: Medium.
- happier #499: retained-message submission fix (merged).
- paper-qa #1302: citation ingestion fix (merged).
