System design — given a model and a set of requirements, how would they design the production system? - Code quality — not whiteboard algorithms, but production-quality code with tests, documentation, and error handling - Debugging — give them a broken ML pipeline and ask them to diagnose and fix it