Went through the Zoom ML engineer interview loop earlier this year for a role on their AI/ML team working on real-time audio and video enhancement. Sharing this because the process is a little different from the standard FAANG MLE loop.
First, a note on scope: Zoom's ML work splits roughly between the applied science side (improving call quality, noise suppression, background blur, VB, lip sync detection) and the more traditional SaaS ML stuff (meeting summaries, transcription, AI Companion features). The role I interviewed for was the former, so your mileage will vary on the latter track.
Loop structure:
Recruiter + HM screen: standard. They asked about background, what kinds of models I've deployed to production, and whether I have experience with latency-constrained inference. That last part matters a lot for the audio/video track.
ML breadth round: a 60-min technical discussion, no code. Topics covered: bias-variance tradeoff, regularization, gradient descent variants, how I'd approach class imbalance in a training set, and one question about evaluating a ranking model. Nothing exotic but they go deep on follow-ups. Know your fundamentals well.
Coding round: two problems. One standard medium LC (I got a graph traversal). One ML-adjacent: given a stream of events, implement an online moving average with some constraints. The second one was more interesting. It's basically testing whether you can write clean numerical code.
ML system design: design an automated noise suppression system for real-time audio. This was the meaty one. They want to see: how you'd frame the problem (supervised vs. unsupervised, what labels you'd collect), model architecture at a high level (they mentioned deep learning context but didn't require it), how you handle inference latency on-device vs. server-side, and how you'd evaluate and monitor quality post-launch. I drew a lot on SNR, PESQ metrics, and A/B testing call quality. Have something concrete to say about latency budgets.
Behavioral: 5 questions, all behavioral, standard STAR. One specifically about a time your model underperformed in production. Have a real story there.
Bar felt senior-mid level. Less algo-heavy than Amazon or Google MLE loops, but the domain knowledge on real-time ML and latency constraints was expected. If you're coming from pure NLP or recsys without any signal processing background, I'd study up on at least the conceptual side of audio ML.