Senior LLM Engineer — AI agents in production, legal-tech (Paris)

The context

A Paris legal-tech is building an AI agent for notaries and real-estate professionals: sale deeds, estates, powers of attorney, emails. The agent is in production and paid for by its customers, with thousands of users every month. The company has raised several million euros; the engineering team is six people and works directly with the founder, who is highly technical.

The 2026 project: turn the agent from an assistant that answers into a much more proactive agent that triggers itself (incoming email, new document, scheduled job), chains calls to many internal and external tools, and produces documents. You will help design that architecture.

What you will do

• Design and evolve the agentic graph: the agent picks its own tools and path, orchestrates many internal and external API calls, and keeps state across long-running tasks.
• Keep production healthy: unstable external APIs, retries, fallbacks, degraded mode when a service goes down.
• Build the quality loop: trace analysis, versioned eval datasets, regression checks in CI, continuous improvement from what breaks.
• Bring the team what you have already learned elsewhere: the known traps and the patterns that hold at scale.

Who you are

• You have already built AND run in production a non-deterministic agentic system, with many external tool calls and real volume (thousands of users or runs). Not a POC, not an internal tool for ten people.
• You can walk through a production run that went wrong: how you detected and understood it, and what you put in place so it would not happen again.
• You have real practice with evals and LLM observability (LangFuse, LangSmith, Braintrust or similar): you catch a quality regression before your users do.
• You are a software engineer first: strong Python (typing, async, tests), architecture, API integrations. A data science background is welcome but not enough: what matters is agentic systems shipped to production.
• Senior individual contributor role, no people management.
• Bonus: knowledge graphs or structured extraction, an interest in law.

The process

A one-hour, fully technical interview with the founder: he takes one system you built and digs to the bottom of it (users, volumes, architecture, your role, incidents, evals). Then a take-home exercise, a code walkthrough with the team, references and the offer.