Scale AI · Primly Community

Scale AI data scientist interview: SQL, case study, and stats. What to actually prep.

de_derek · 6 replies

Finished my Scale AI data scientist interview loop about three weeks ago. Sharing the breakdown because I couldn't find much when I was searching.

The loop was four rounds total. Recruiter screen, then three back-to-back on separate days.

SQL round. This was tougher than the average DS SQL screen. Not just basic aggregations. They gave me a schema that modeled something close to their tasker annotation pipeline and asked me to find quality outliers. Think: which taskers have an acceptance rate below a threshold AND have done more than N tasks in the past 30 days. Window functions came up. I stumbled on a PARTITION BY question initially and they let me recover, which was appreciated. Expected level: you should be comfortable with CTEs, window functions, and being able to explain WHY you'd use a particular join type.

Case study. This was the most interesting round. They gave a scenario where annotation quality is degrading on a specific task type (image segmentation, in my case). How do you detect it, how do you measure it, how do you fix it. They want you to think in terms of data pipelines, not just notebook analysis. I framed my answer around: detection (statistical control charts, p-charts for defect rates), diagnosis (cohort the taskers by tenure, device type, time of day), and intervention (targeted review queue, hold-out gold sets for calibration). They pushed back on my calibration idea and we had a real back-and-forth. That felt intentional.

Stats and experiment design. Fairly standard A/B testing setup but with a twist: low sample sizes. The question was basically, when should you stop an experiment early when your population is small. I talked through Bayesian approaches vs. sequential testing and they seemed to like that I knew both. They did ask one probability question that felt more like a brain teaser than practical stats. Don't overthink those.

Overall. The loop felt substantive. They're not just going through motions. The people I spoke with clearly knew the domain. Comp offer for a mid-senior DS role in SF was around $195k base with equity on top, 2026. I didn't get a detailed equity breakdown before accepting so I wish I'd pushed on that earlier.

Happy to answer questions.

6 replies

analyst_ana

This is really helpful. Did they tell you ahead of time that SQL would be live coding or was it more of a discussion? I always panic when I can't see whether my query actually runs.

consultant_cam

Live coding in a shared editor. Not a real database behind it, so you can't run it, just write it and talk through it. They're fine with minor syntax issues as long as the logic is right. I messed up a window function alias and just said 'I'd fix that in a real environment' and moved on.

alex_design

$195k base SF mid-senior DS 2026. Equity: did they give a number at all or just 'equity TBD'? Trying to build comp data for this company.

ux_uma

They gave a range during recruiter call. Said $180-220k base depending on level. Equity was presented as a 4-year RSU grant amount in dollars but they didn't break it down by vesting cliff until the offer letter. It was a standard 1-year cliff, then monthly.

returner_ren

Good write-up. The annotation quality case is actually really on-brand for Scale. They care a lot about data quality internally. Makes sense they'd probe that. Did the case feel like it required Scale-specific knowledge or could you have answered it from general DS background?

firsttime_mgr

Mostly general DS background. You should understand what RLHF pipelines broadly look like and why data quality matters at that stage, but they gave enough context in the prompt itself. I'd read their blog before the loop just to get the vocabulary.