Mistral AI Interview Process - Senior ML Infrastructure Engineer

I interviewed with Mistral AI in London in August–September 2026. The process took about five weeks and was entirely virtual. I had ten years’ experience, mainly in backend infrastructure and distributed systems.

Opening conversations

The recruiter asked why I wanted to support researchers and how much code I still wrote. In the project discussion, we dug into a dataset cache I’d built, especially when reusing a result was safe. An undeclared environment variable exposed a gap in my explanation.

Python coding

Merge timestamp-sorted logs from several files without loading everything into memory. I used a min-heap, then got the follow-up: add checkpoint and resume.

I initially suggested saving reader offsets, forgetting the records already buffered in the heap. Restarting from those offsets would skip records. I changed it to save the next unreturned position for each source. Checkpointing before the first next() call caught the bug.

Technical quiz

Questions covered why async hadn’t improved a pipeline, timing GPU operations, GPU memory usage, and what a training checkpoint needs besides model weights. The interviewer kept following up on my answers. I was less sure about investigating allocations in an unfamiliar custom operation.

Code review

I reviewed an async evaluation client. Failed requests were excluded from the score, the cache could reuse responses across model versions, and the code created too many tasks upfront.

I spent too long commenting on names before spotting the scoring problem. That was the round where I felt most rushed.

System design

Design a service to evaluate model checkpoints on shared GPUs. We discussed incomplete uploads, two workers finishing the same task, and whether new checkpoints should replace older queued evaluations.

I proposed immutable run specifications, a manifest marking completed uploads, and separate outputs for each task attempt. This was my strongest round.

Collaboration

A researcher thinks your safeguards make experiments harder. How do you respond? I talked about relaxing checks for isolated experiments while keeping stronger checks for shared infrastructure.

My advice: practise practical Python tasks, including testing and recovery. In code review, start with anything that could make the result wrong.