just finished my amazon DS loop last month, applied for a science role on the advertising team in seattle. sharing what i found because there's surprisingly little specific info out there on the DS track vs the SWE track.
five rounds total, one recruiter screen and four interviews. the rounds were:
round 1: SQL + product sense. they gave me a schema with 3 tables (users, events, orders) and asked me to write a query for weekly retention cohorts. nothing exotic but it needed a self-join. then a product question about a metric that dropped 15% week over week. they wanted you to walk through a structured diagnosis, not just list possible causes. i think this is where a lot of DS candidates get cut. they ramble. be systematic.
round 2: stats / experimentation. one question on AB test design, power analysis, minimum detectable effect. they asked what you'd do if your experiment had a novelty effect. another question was about p-hacking and how you'd prevent it in a high-velocity testing culture. this felt very amazon-specific. they run a lot of experiments and they want people who won't just ship false positives.
round 3: ML applied. describe a model you built from scratch. i walked through a churn model i built at my current job. they pushed hard on feature engineering, how i handled class imbalance, and what happened when it went to prod vs staging. they don't want a textbook answer. they want a story where something went sideways.
round 4: leadership principles. this is the behavioral round and it matters A LOT at amazon. like, more than most companies. i prepped about 12 LP stories using STAR. they asked about customer obsession, dive deep, and deliver results specifically. they probed for specifics hard. if you say 'we increased revenue', they want the number.
a few things i noticed: the interviewers all asked 'what would you do differently?' at the end of STAR stories. build that into your prep. also, no take-home in my loop but i've heard other teams do one.
i got an offer. L5 DS, $190k base + RSUs + sign-on but i'll save the full breakdown for the comp thread.