Went through the full Apple MLE loop in February for a role on the Siri/Foundation Models team. ICT5 level. Sharing because ML interview prep advice is all over the place and Apple specifically has a different flavor from Google/Meta.
Loop structure. Six rounds total: recruiter screen, ML phone screen, then four onsite rounds. The onsite breakdown: ML theory/fundamentals, ML system design, coding, behavioral.
ML phone screen. One hour with a senior MLE. Half probability/stats (Bayes theorem, bias-variance tradeoff, what happens when you over-regularize), half ML concepts (how does gradient descent actually work, explain attention from first principles). Not trivia-hunting. They want to see you reason through things. I got asked to derive the softmax gradient by hand. That's probably team-specific so don't panic if yours is different.
ML theory round. Deep. We spent 25 minutes on transformer architecture (attention mechanisms, positional encoding, why layer norm before attention vs after). Then a probability problem: given a sequence of biased coin flips, derive a maximum likelihood estimate for the bias. They explicitly said calculators were fine, they wanted to see the reasoning.
ML system design. "Design a personalized notification ranking system for Apple devices." This one is really about tradeoffs: on-device inference vs server-side, privacy constraints (Apple is obsessive about this), latency budgets, how you'd handle cold-start for new users. The privacy angle is not decoration. They specifically probed what data I'd use and whether it would leave the device. Have a real answer about federated learning or differential privacy.
Coding round. Medium-hard LC difficulty. I got a graph problem. Not ML-specific, just standard DS&A. Don't underestimate this round.
What was different from other MLE interviews I've done. The privacy angle is genuinely central, not a checkbox. They also cared a lot about how I'd evaluate models in production, not just offline. Online A/B test design, handling distribution shift, knowing when a model is degrading before users complain. Those topics came up in two different rounds.
Total time from first contact to verbal offer: 11 weeks.