Back-of-envelope estimation, ML vs non-ML decisions, NFR trade-offs, and problem framing — the two cornerstones every ML system design interview starts with. Includes classic phrasing traps and scenario-based reasoning.

Lambda vs Kappa, exactly-once semantics, schema evolution, temporal leakage, high-cardinality encoding, and embedding strategies at scale. Tests pipeline reasoning and the subtle traps that break production feature systems.

Online/offline store consistency, point-in-time correctness, distributed training strategies (data vs model parallelism), ZeRO optimizer stages, and GPU utilization bottlenecks. Covers the storage-to-training pipeline that separates ML practitioners from ML engineers.

Blue-green vs canary deployments, shadow mode, multi-armed bandit routing, batch vs real-time inference cost trade-offs, and hybrid architectures. Tests the production deployment reasoning that interviews focus on most.

A/B test validity, SUTVA violations, novelty effect, SRM detection, feature drift monitoring, data quality alerting, and feedback loop identification. Two of the most misunderstood ML system components — together in one test.

Two-tower models, popularity bias, diversity-relevance trade-offs, BM25 vs semantic search, ANN indexes (HNSW), and cascade ranking. Two of the most common ML system design interview topics — side by side.

KV cache management, continuous batching, PagedAttention, chunking strategies, retrieval quality vs hallucination trade-offs, and hybrid dense-sparse retrieval. The two hottest ML system design topics in modern interviews.

Agent orchestration, prompt injection attack surfaces, guard-rail design, Pareto frontier optimization, latency vs accuracy trade-offs, and the irreversibility of production ML decisions. Caps the curriculum with the hardest design thinking in modern ML.

Simulates an entry-level ML system design interview round. 10 scenario-based questions covering all 16 topics — each with a classic reasoning trap. Great for first mock attempt or knowledge check before a FAANG phone screen.

Second easy mock — different questions, same difficulty ceiling. Tests the complementary half of the fundamentals not covered in Mock 1. Includes topics 03, 05, 06, 08, 12, 14 to complete the easy-level tour of all 16 topics.

Simulates a mid-level ML engineering interview. 12 scenario-based questions requiring multi-step reasoning across pipeline design, training infrastructure, experimentation, LLM serving, and GenAI patterns. Each question contains a deceptive option that trips up candidates with 2–3 years of experience.

Second medium mock — a fresh set of 12 questions emphasizing engineering trade-offs: cost vs latency, batch vs real-time, canary deployment criteria, and monitoring for ML-specific failure modes. Complementary to Mock 3 with no repeated questions.

Simulates a senior ML engineer interview loop. 15 hard scenario questions spanning queuing theory at scale, distributed training internals, KV cache thrashing, multi-hop RAG failure modes, and production feedback loops. Every question is a FAANG-caliber trap with non-obvious answers.

Second hard mock — 15 different questions at the same difficulty ceiling. Emphasizes ZeRO optimizer stages, HNSW graph degradation, agentic loop prevention, gRPC routing pitfalls, and statistical validity in experimentation. No repeated questions from Mock 5.

18 of the hardest ML system design questions in the bank. Covers all 16 topics at maximum depth. Questions test multi-step architecture reasoning, production failure analysis, and the non-intuitive behavior of distributed ML systems. Intended for senior/staff engineers preparing for AI architect screening or ML platform tech lead rounds.

18 elite-level questions with zero overlap from Elite 1. Focuses on production failure modes: causal framing failures, Flink state thrashing, TOCTOU race conditions in async LLM pipelines, Thompson Sampling incompatibility with two-tower models, and PPV/NPV at population scale. Combines questions from the hard practice bank with the hardest questions from the per-topic files — the final boss of the ML system design curriculum.