I build AI agents that run in production, for real clients, with real money moving through them.
Based in Guadeloupe, French Caribbean. That puts me at UTC-4, the same working hours as New York and Boston.
A sports club, fully digitised. BTB 94 runs its 165 licensed members, 7 teams, online payments and a granular permission model on a system I built and still operate. When the federation updates a licence, a scheduled job syncs it. When something breaks at 3am, it comes back to me.
An agent platform in production. 123 serverless functions with strict modular separation, a tool-calling reasoning core routing over a registry of typed tools, and a 10-agent voice team doing live agent-to-agent handoffs while preserving conversation state. I migrated the whole thing across model providers by diffing behaviour function by function.
MCP servers, published. JSON-RPC over Streamable HTTP, 50+ tools exposed to LLM clients, listed in the official Model Context Protocol registry. I wrote one from scratch for a self-hosted CRM because the off-the-shelf integration did not exist.
Field tools that people use on site. A PWA for a safety coordinator who needs it to work offline, standing in a half-built structure with no signal. An equipment rental platform handling fleet, quotes, scheduling and invoicing. Neither is glamorous. Both get opened every working day.
Sole engineer on about 15 production systems since 2020. Scoping, build, deploy, monitoring, incident response. There is no other team to hand off to, so every reliability failure lands on my desk. That shapes how I write things.
I run a security audit practice alongside the engineering work: static analysis, access control testing, privilege escalation, signed webhooks. I test my own systems the same way, which is how I found an SSO layer quietly returning 302 to an automated job instead of running it.
Mostly TypeScript and Python, with SQL underneath. Node, Deno, Next.js, PostgreSQL, Docker, serverless. On the model side: Claude and OpenAI APIs, MCP, RAG, multi-turn state, trajectory debugging. And Playwright, usually against targets that would rather not be automated.
portfolio-laurent.vercel.app · LinkedIn
Open to remote engineering work and consulting. French is my first language. I work in English every day, mostly in technical documentation and model specs.