Business constraints first: Millisecond latency (online inference often under 10-50ms), massive request volume, and scores that go directly into bidding - so calibration matters as much as ranking.
High-dimensional sparse features: User / ad / context ID features become embeddings. Explain feature crossing evolution from FM to DeepFM/DIN and what each generation solves.
Calibration: Model output must map to true CTR (Platt scaling / isotonic / per-segment calibration), otherwise the bidding system systematically overpays or underpays.
Freshness: Streaming sample joins (impression-click join with click delay and attribution windows), near-real-time online learning for distribution drift, and probability correction after negative downsampling.
Get asked follow-ups live, then receive a scored report — like a real MLE interview loop.
Start AI mock interview