Design ML systems the way interviewers ask.

Return visually similar images for an uploaded or cropped photo using learned embeddings and approximate nearest neighbour search.

Detect and blur every face and licence plate across billions of street-level images, with a human review loop for ambiguous cases.

Retrieve and rank relevant videos for short, ambiguous queries using text, visual and audio signals across billions of videos.

Catch hate speech, violence and spam across text, images and video with a cascade of models, a policy engine and human moderators.

Recommend videos a user will watch next by combining candidate generation from embeddings with a watch-time ranking model.

Suggest upcoming events a user is likely to attend using location, time and social signals, while handling cold start.

Predict the click probability for each ad in milliseconds, then apply budget, frequency and diversity rules before serving.

Show listings similar to the one a guest is viewing, using learned listing embeddings and real-time availability filters.

Rank posts and articles by predicted engagement, then re-rank for freshness, diversity and content moderation.