Lead LLM

Licorne Society a été missionné par une startup IA en pleine croissance pour les aider à trouver leur Lead LLM Engineer.
What you will own

You will be responsible for one thing:
Make our AI outputs reliable, fast, and indispensable in real workflows.
Concretely:

• Design and evolve our LLM / agent architecture
• Own output quality across key use cases (emails, document analysis, etc.)
• Build evaluation systems (datasets, metrics, regression detection)
• Drive fast iteration loops from production data
• Improve retrieval, reasoning, and tool usage
• Ensure production reliability (latency, failure modes, fallback)
• Work directly with product + founders on what to build and why

What this role is really about

Most teams fail because:

• they don't know what "good output" means
• they don't have evals
• they iterate randomly
• they overuse agents

Your job is to fix that.
You will turn:

• vague user problems
• → into structured AI systems
• → with measurable performance
• → that improve every week

What you need to be excellent at
1. Shipping real LLM systems

• You've built systems used in production (not demos)
• You understand RAG, tools, agents, structured outputs
• You can design full pipelines, not just prompts

2. Evaluation-driven development

• You know how to define quality metrics
• You build datasets from real usage
• You run continuous evals to prevent regressions

3. Debugging complex failures

• You can trace issues across:
• retrieval
• prompts
• model behavior
• You don't guess - you isolate and fix

4. Speed of iteration

• You move from problem → improvement in hours or days, not weeks
• You use logs, traces, and data - not intuition alone

5. Strong judgment

• You know when to:
• use an agent vs a pipeline
• add complexity vs simplify
• You optimize for reliability and user value, not novelty

What we don't care about

• Number of years of experience
• Whether you've used a specific framework
• Fancy research credentials

If you can build, debug, and improve real systems, you're a fit.
What success looks like (first 90 days)

• Clear eval framework for core use cases
• Measurable improvement in output quality
• Faster iteration cycles across the team
• Reduced hallucinations / failures
• Stronger system architecture decisions

Stack (context, not requirements)

• Python (FastAPI)
• Postgres
• Google Cloud
• LangGraph / LangChain (evolving)
• PostHog (product analytics)
• Langfuse (LLM traces)
• LLM APIs (Azure OpenAI)