Machine Learning Engineer System Design Interview Questions

40 system design questions put to Machine Learning Engineers, 7 of them at Meta and 5 each at Anthropic and OpenAI, with Apple, DoorDash, Amazon, Mercor and Jane Street behind them. The brief is almost never a plain web service: you are designing around a model, so the board fills with candidate generation feeding a ranking tier, embedding indexes and approximate nearest-neighbour search, a feature store and the training/serving skew it is there to prevent, batching against tail latency on scarce GPUs, shadow deploys and rollback, and how drift gets caught once traffic moves. Reported prompts include a scalable recommendation serving system, near-duplicate video detection over a large catalogue, and a harmful-content detector for weapon ads, where the real argument is precision against recall and what the human review queue costs per day. The newer questions show who is hiring: an NL-to-SQL optimisation assistant and an entitlement-aware agentic portfolio workflow, both turning on tool calls, permission checks and offline evaluation rather than model architecture. 23 of the 40 were asked onsite and 15 are rated hard, so expect the follow-ups to push into capacity and cost. All 40 carry a written solution, and 65% are readable without premium.