Walk me through your depth in your ML specialty, and the hard problems in it.

The screen is genuine hands-on depth in one area versus breadth that's all paper-reading; the tell is whether you talk about failure modes and operational pain (the things only people who shipped know) rather than benchmark leaderboards. The held-back follow-up is a sharp drill into a specific sub-problem you named, so claim only depth you can survive five questions about. Pick ONE specialty and go deep; a candidate who claims expert depth in NLP, vision, recsys, and RL all at once is signaling the opposite.