Machine Learning System Design Interview
For software and machine learning engineers preparing for the ML system design interview. You will learn a seven-step framework and practise it on ten case studies, from visual search and content moderation to recommendations, ad click prediction and People You May Know, so you can turn a vague product prompt into a design you can defend. 33 lessons in 11 sections, about 26 hours.
Course Content
Learn the seven-step framework — framing, metrics, data, model, training, serving and monitoring — that every case study in this course follows. 3 lessons, about 180 minutes.
Design image-based product search by training image embeddings contrastively and serving them with approximate nearest neighbour search. 3 lessons, about 130 minutes.
Design a batch pipeline that detects and blurs faces and licence plates in street imagery, where a missed detection is a privacy incident and recall dominates. 3 lessons, about 120 minutes.
Design video search that combines lexical and semantic retrieval with visual content understanding, then ranks the results within a tight latency budget. 3 lessons, about 130 minutes.
Design a multimodal moderation system that catches rare harms, sets per-category thresholds under a real cost asymmetry and runs as a tiered pipeline with human review. 3 lessons, about 140 minutes.
Build the canonical two-stage recommender — candidate generation, then multi-objective ranking — and handle feedback loops, exploration and user wellbeing. 3 lessons, about 140 minutes.
Recommend local events when every item is new, expires on a fixed date and is limited by distance, using content features, hard filters and gradient-boosted trees. 3 lessons, about 130 minutes.
Predict ad clicks with probabilities calibrated well enough to set prices, using very sparse features at scale and scoring inside a real-time auction. 3 lessons, about 140 minutes.
Learn listing embeddings from browsing sessions, adapt them for bookings, markets and new listings, and serve similar listings from precomputed neighbours. 3 lessons, about 110 minutes.
Rank a personalised feed by predicting several outcomes at once and combining them into one score that encodes product policy. 3 lessons, about 130 minutes.
Suggest connections on a social graph with graph features, two-hop candidate generation and tree models, while guarding against unwanted contact and privacy leaks. 3 lessons, about 120 minutes.