The ML system design framework

ML system design interviews are deliberately vague: 'Design YouTube recommendations' or 'Build a fraud detector'. There is no single right answer. Interviewers grade how you structure the problem, which tradeoffs you notice, and whether your design could actually run in production.

A fixed framework stops you from jumping straight to 'I'd use a transformer'. Spend the first 5–8 minutes on requirements — that is where most candidates lose points.

1. Clarify requirements   users, scale, latency, constraints
2. Frame as an ML task     input → output, objective, labels
3. Data                    sources, labels, features
4. Model                   baseline first, then improve
5. Evaluation              offline metrics + online A/B test
6. Serving                 latency, throughput, infrastructure
7. Monitoring & iteration  drift, retraining, feedback loops

Good clarifying questions: How many users and items? What latency is acceptable? Is it real-time or can it be precomputed daily? What does success look like for the business? Do we have labels, or do we need to collect them? Are there privacy or fairness constraints?

Always propose a simple baseline before anything deep: popularity ranking, logistic regression, or a hand-written rule. It gives you a benchmark, ships quickly, and shows maturity.