Everyone does Project 1, then picks one Project 2 (Science or Systems). Same credit either way; pick by interest, not difficulty.
The big picture for each project. Click a title for the full spec.
The whole project in one flow: pick a big public dataset, ask three decision-worthy questions, answer them with real SQL (joins, CTEs, subqueries, window functions), and hand-trace two debug tables to prove you know how it runs.
Project 1 was SQL on one warehouse. Project 2 adds the three layers a real data team owns: a Spark job to ingest outside data, embeddings for semantic search, and a BigQuery ML model that makes a calibrated business call.
You build the whole engine from scratch: an inverted index with BM25 for keywords, LSH over vectors for meaning, a hybrid layer that fuses them, and a planner that picks the plan, then race it for parity against the production tools shipping in a billion phones.
The full AI Policy lives on the course logistics page. Read it here.