Just cleared the Vercel ML interview loop (declined the offer, different story, not going to editorialize). Figured I'd write it up since the ML-specific Vercel content is basically nonexistent.
Context: the role was on their infrastructure intelligence team, basically ML to improve build performance, predictive caching, that kind of thing. Not a GenAI or LLM product team. Keep that framing in mind because it shapes what they ask.
Recruiter screen: Standard, 30 minutes. They were honest that the ML team at Vercel is small and this hire would be doing both modeling and some production engineering. They explicitly asked about my comfort with production systems, not just notebooks.
Technical screen: Sixty minutes. One coding problem that was Leetcode medium territory, then a conceptual ML discussion. The discussion was good. The interviewer asked me to walk through how I'd approach a problem where build time prediction has sparse historical data (cold start problem). Not about any specific framework, more about your intuition around handling sparsity: fallback heuristics, feature engineering, when you'd use a simpler model vs. a more complex one. They were not impressed by the person who names the biggest model and calls it done.
ML system design round: This was the real test. Build a system that decides which cached build artifacts to invalidate. Think about this: it's not a standard recsys problem, it has constraints around latency (the decision has to be fast) and data freshness. They probed hard on evaluation: how do you measure whether your model is actually helping build times, what's your offline eval strategy, how does it degrade gracefully if the model is wrong.
Coding round: More standard, two problems. Trees and graph traversal. Nothing ML-specific.
Behavioral: One round, mix of classic behavioral questions and motivation questions. They asked why I'm interested in infrastructure rather than product ML, which is the right question honestly.
Overall the loop felt like they want a real ML engineer who can ship, not a researcher. If your background is 80% Kaggle and 20% prod, you'll feel that gap. If you've actually put models in services, you'll be fine.
Timeline was about 6 weeks from first call to offer.