Google · Primly Community

Google machine learning engineer interview: what rounds look like and where people fail

ml_mike · 4 replies

I went through the Google MLE interview for an L5 role, Mountain View, in Q1 2026. The team was in the Ads space. Sharing this because I had to piece together a picture from 4-year-old posts and the process has shifted.

The loop was 5 rounds:

Coding (x2): Standard SWE algo. They don't give you ML-specific problems in the coding rounds. One was a classic heap problem, one was a DP problem that I thought I was going to botch but recovered. Google MLE coding difficulty is the same as SWE. If you're rusty because you've been heads-down in PyTorch for two years, get on Neetcode immediately. This is not optional.

ML Design: Designing a click-through rate prediction system from scratch. Covered: problem framing (binary classification), feature engineering for sparse categorical features (user/item IDs), model choice (logistic regression vs. GBDT vs. neural), training pipeline, offline vs. online evaluation, and deployment. They went deep on calibration of probability estimates, which was a good sign the interviewer knew their stuff. For Ads-focused roles, know the basics of how ad auction mechanics interact with CTR models.

ML Theory: Mix of ML fundamentals. Bias-variance tradeoff, gradient descent variants, what happens when your training set has class imbalance, precision vs. recall tradeoffs. Then a question about how you'd debug a model that performs well offline but degrades fast in production (concept drift). That last question is where I've seen people stall. Have a mental framework for diagnosing model degradation.

Behavioral: Googleyness. Conflict, leadership, ambiguity. Nothing MLE-specific. Bring strong STAR stories.

The interview took about 6 weeks from recruiter screen to offer. The ML design round was definitely where I felt most evaluated. TC came in around $380k at L5 for Mountain View. I expected more given market comps I'd seen online but the team and scope felt right.

Feel free to ask about specific rounds. The ML design one had a lot of ground to cover.

4 replies

ds_dmitri

The concept drift debugging question is something I've seen at multiple companies lately. Do you have a clean framework for it that worked in the room?

ml_mike

Yeah. I structured it as: first diagnose whether it's input drift (features shifted), label drift (outcome distribution changed), or feedback loop drift (your model's predictions are influencing future training data). Then walk through monitoring strategy for each. The interviewer lit up at feedback loop drift because that's the interesting real-world case for an ads model.

corp_refugee

$380k at L5 MLE for MTV in 2026 feels a bit low based on what I've seen but I also don't have a large recent sample. Anyone else have recent data points?

contractor_kai

I keep seeing the 'coding rounds are the same as SWE' warning from everyone who goes through Google MLE. And yet ML candidates keep underpreparing for it. Two separate friends failed the MLE loop purely on coding. The ML parts were fine. Just do the grind.