Research Engineer, ML Eval Platform: Scale & Trust

Mistral.ai; Paris; On-site; EUR 85,000 - 135,000; Full time

Job summary

Mistral.ai in France is seeking a Research Engineer for the Eval Platform to build the infrastructure needed to measure model quality, ensuring evaluations are reproducible, scalable, and fast across GPU clusters.

You will collaborate with researchers to turn evaluation needs into robust tooling, enabling accessible results via APIs and dashboards and supporting agentic, multi-turn and tool-using models.

Qualifications

• Master's or PhD in Computer Science, or equivalent experience.
• 4+ years building production-grade software, ideally large-scale ML codebases or distributed systems.
• Excellent Python and strong software-design instincts: testing, code review, CI/CD.
• Experience running workloads on GPU clusters (Slurm, Kubernetes, Ray or similar).
• Familiarity with LLM inference and evaluation.
• A product mindset: researchers are your users.
• Self-starter, low-ego, collaborative.

Responsibilities

• Build systems that keep eval results reproducible and comparable over time, as models, benchmarks and code evolve.
• Run evaluations at scale across our GPU clusters, from model serving to scoring.
• Make eval results easy to access, explore and trust, through APIs and dashboards that researchers use every day.
• Catch broken or noisy evals before they mislead research decisions.
• Support evaluation of agentic, multi-turn and tool-using models.
• Work closely with researchers to turn new evaluation needs into robust, shared tooling.

Skills

Python

Distributed systems

Testing & code review

CI/CD

Education

Master's or PhD in Computer Science or equivalent

Tools

Slurm

Kubernetes

Ray

vLLM

SGLang

Job description

Mistral.ai in France is seeking a Research Engineer for the Eval Platform to build the infrastructure needed to measure model quality, ensuring evaluations are reproducible, scalable, and fast across GPU clusters.

You will collaborate with researchers to turn evaluation needs into robust tooling, enabling accessible results via APIs and dashboards and supporting agentic, multi-turn and tool-using models.