Design production ML systems end to end: problem framing, training data, features, evaluation, serving, and everything that breaks after launch.
What you'll learn
Turn a vague product ask into an ML problem with clear objectives, labels, and success metrics.
Design training data, features, and embeddings that hold up outside the notebook.
Choose models and evaluation strategies that match the business goal, not just offline accuracy.
Serve, monitor, and retrain under drift, latency budgets, and real traffic.
About this track
ML system design interviews are not about naming a model. They ask whether you can take a fuzzy product goal, turn it into a learning problem, and reason about the pipeline that keeps it working in production.
This track walks the full loop. You frame the problem and choose labels, design training data and features, pick a model with generalization in mind, and select evaluation metrics that track the business goal rather than the loss curve.
Then you apply it. Each question follows a real interview arc — recommendations, ranking, fraud, harmful content, ETA prediction, retrieval-augmented assistants — including the serving path, the monitoring story, and what happens when the data drifts.