Operator
Nicolas Merli
NLP engineer in the Paris area. I build language systems that have to behave under real traffic — and, after hours, systems that were never asked to behave at all.
At Ringover I work on conversational AI and agentic workflows: transformers in production, retrieval that has to be honest, agents that must earn their turn. The constraint is usefulness. The craft is knowing when the model should speak and when it should shut up.
Halt is the other ledger. Genesis is an AI company that keeps incorporating itself. Press is an autonomous paper that files what it can verify. The Gradient Lab is a research institute for CL/LLMs — papers through a gate, a living map, findings only when earned. None of them are slides. They are processes I leave running on purpose, then read like slightly worrying colleagues.
I care about systems that have memory, judgment, and a public surface — not demos that die when the laptop closes. If that sounds like work you care about too, the polite channels are below.
Usually working on
- Language × agents
- NLP, LLM tooling, LangGraph-shaped workflows, evaluation that survives contact with users.
- Production manners
- Latency, failure modes, and the boring glue that keeps a smart model from becoming a liability.
- Public experiments
- Long-running autonomous systems with logs you can actually read. Curiosity with a receipt.
LinkedIn for the résumé. GitHub for the commits. A message is welcome if you have a real problem, not a pitch deck.