Amazon · Primly Community

Amazon machine learning engineer interview: what they actually ask vs. what you'd expect

ml_mike · 4 replies

going to give you the real version of the amazon MLE loop, not the polished one. i just finished it for an L5 role on the personalization team. six rounds including a phone screen.

phone screen: one coding problem (medium lc, tree traversal) and two conceptual ML questions. the ML questions were: 1) how do you evaluate a ranking model? 2) what's the difference between bias and variance and what would cause each in practice? if you've done any real ML work these aren't hard but they screen out people who memorized terms without understanding them.

onsite round 1: coding. two coding problems. medium difficulty, arrays and graphs. nothing ML-specific. this is where people expecting a pure ML interview get burned. amazon MLE interviews have significant SWE content. you need to pass the coding bar, period.

onsite round 2: ML systems design. design a recommendation system for amazon product search. this is where depth matters. they don't want textbook collaborative filtering. they want: how do you handle cold start? what features do you pull? how do you balance exploration vs exploitation? how do you set up your offline eval and then your online AB test? i talked about candidate generation then ranking as a two-stage approach and they seemed very familiar with that framing (probably because that's how their systems actually work).

onsite round 3: applied ML. walk through a real model you built end to end. same as DS loops. they want the prod story including failure. i talked about a recsys i built that worked great offline and was mediocre online and we figured out it was leaking future information during feature engineering. that kind of war story lands better than a success story.

onsite round 4: ML fundamentals. deep questions on things like: gradient descent variants and when you'd use each, regularization, handling imbalanced classes, when you'd use tree-based vs neural approaches. more theoretical than the applied round.

LP round: deliver results and think big. have actual impact numbers.

honest take on the loop: it's rigorous and fair. i didn't feel like i was fighting arbitrary gotcha questions. the SWE component is real though. if your coding is weak you won't clear the bar even if your ML depth is strong. fix that first.

offer came through, l5, AWS AI services team base around $195k + RSUs, details approximate.

4 replies

ds_dmitri

the future leakage story in the applied round is such a good move. every serious ML team has been burned by that at some point and it signals you've actually shipped something to prod.

quietquit_quincy

how hard was the coding bar relative to an SWE loop at amazon? like was it the same lc difficulty or lighter because the ML rounds are supposed to be the real signal?

ml_mike

honestly comparable difficulty. i'd say medium lc is the baseline for both rounds. they don't give you a pass on coding because you're going for MLE. from what i heard in the debrief call, coding is a hard filter and ML depth is what differentiates above the threshold.

recruiter_rita

the two-stage candidate generation then ranking framing is something amazon ML teams respond well to because it maps to how they actually build. knowing the vocabulary of the domain you're interviewing into always helps, not because you're performing but because you're speaking the right language.