Mistral AI · Primly Community

Mistral AI data scientist interview (SQL + case + stats), my full loop breakdown

analyst_ana · 6 replies

Went through Mistral's DS loop in early 2026 for a role on their inference analytics team. Applied through LinkedIn, heard back in about 10 days. Want to share because when I was prepping I found almost nothing useful about their DS-specific process.

The recruiter call was pretty standard. Maybe 25 minutes, covered background, why Mistral, rough timeline. She was direct about the fact that it's a small team and they move fast when they want someone.

Phone screen with a DS on the team This was 45 minutes, half technical half fit. SQL was the first thing that came up. They gave me a prompt about analyzing model usage data: write a query to find users whose API call volume dropped more than 40% week over week for at least two consecutive weeks. Involves window functions, CTEs, the usual. Nothing tricky but you have to actually know SQL, not just vaguely remember it.

The second part was a stats question. They asked about how you'd detect if a new model version was performing differently on a long-tail of prompts vs the main distribution. They wanted to hear about stratified sampling, not just overall averages. I mentioned Mann-Whitney U and they seemed happy I didn't immediately jump to t-test.

The case round This was the most interesting one. They framed it as: Mistral is launching a new API pricing tier. How do you measure whether it's cannibalizing the higher-margin tier vs growing the overall market? Walk us through your analysis plan.

They're specifically checking whether you think about causal inference vs pure correlation, what data you'd need, and how you'd communicate findings to a non-DS stakeholder. I fumbled a bit on the stakeholder communication piece. Something to prep.

Onsite (4 rounds, Paris remote via Zoom) Product sense, a deeper stats round (A/B testing at scale, when to use Bayesian vs frequentist approaches for low-traffic segments), a system design-light round about how you'd build a real-time eval pipeline for LLM outputs, and behavioral.

The behavioral round felt genuinely curious, not formulaic. They asked about a time I pushed back on a stakeholder's interpretation of data. They wanted specifics, not a cleaned-up story.

Total time from first contact to offer: about 5.5 weeks. Offer came on a Friday.

6 replies

analyst_ana

The consecutive-week drop question is sneaky. Did they expect you to handle edge cases like users who had zero calls one week? Like NULL vs 0 matters a lot there.

ds_dmitri

Yes, and they actually asked me about that explicitly after I wrote the initial query. I'd coalesce NULLs to 0 and they wanted to know if that was the right call or if zero-call weeks should be excluded from the consecutive-drop check entirely. We went back and forth on it for like 5 minutes. No single right answer but they wanted to see you reason through it.

ml_mike

The eval pipeline round is interesting. Mistral is genuinely thinking hard about automated evals for open-weight models. Did they ask about latency constraints or was it more about correctness/quality metrics?

ux_uma

The stakeholder communication piece tripping you up is so relatable. A lot of DS interviews screen for raw technical skill and then wonder why hires can't communicate findings. Sounds like Mistral at least tries to catch it.

alex_design

Do you know what level this was for? Was it a senior DS role or more mid-level? Trying to calibrate whether the stats depth is the same across levels.

ds_dmitri

Mid-senior, something like 4-6 YOE equivalent. I'd guess a junior DS loop has less of the causal inference stuff and more pure SQL/Python. But I genuinely don't know, they only posted one DS req when I applied.