ML system design
| # | File | Covers |
|---|---|---|
| 01 | ML system design framework | the structure to apply to any question; capacity arithmetic; feedback loops |
| 02 | Design: a RAG assistant over company documents | the most likely AI design question for a backend engineer |
| 03 | Design: a fraud detection system | the canonical classical-ML design: imbalance, cost thresholds, label delay, adversarial drift |
| 04 | Design an agent platform | the multi-tenant platform question: gateway, tool auth, evals, isolation |
| 05 | Design multi-tenant vector search | isolation vs ANN recall; partitioning; where the leak really is |
| 06 | Cut LLM spend without losing quality | the cost-reduction brief, and proving the cheap system is not worse |
| 07 | Design an internal automation platform | ops automation for the business: the glue/rules boundary, ownership, shadow-running |
The two things that carry most questions
Say the outline before diving in. Clarify the decision, frame the ML problem, state the baseline, define metrics, then data, model, serving, rollout, monitoring, failure modes. Candidates who start at model architecture have skipped everything that matters.
Retrieve-then-rank covers search, recommendations, ads and RAG. The capacity arithmetic forces it: scoring every candidate with an expensive model at production QPS is impossible, so a cheap high-recall stage feeds an expensive high-precision one.