Framework · 10 case studies · senior rounds
ML System Design Interview Guide
The round that decides mid & senior offers — and the one with the least free material. A repeatable framework plus 10 worked case studies: recommendations, ranking, fraud, search, RAG, churn, ETA, moderation, and forecasting.
The 8-step framework (use it for any prompt)
-
1
Clarify the problem
Pin down the business goal, the user, scale (QPS, users, items), latency budget, and what 'success' means before touching models. Restate it back to the interviewer.
-
2
Frame as an ML task
Map the business goal to a concrete ML formulation — classification, ranking, regression, retrieval, generation — and define the label. The hardest part is often where the label comes from.
-
3
Pick the metric
Separate the offline metric (AUC, NDCG, RMSE, recall@k) from the online/business metric (CTR, revenue, retention) and name the gap between them. Always propose an A/B test.
-
4
Data & labels
Where does training data come from? Implicit (clicks) vs explicit (ratings) labels, class imbalance, label delay, and feedback loops / position bias.
-
5
Features
User, item, context, and cross features. Discuss freshness (batch vs real-time), an embedding/feature store, and avoiding train-serve skew.
-
6
Model
Start simple (logistic regression / GBT baseline), then justify going deeper (two-tower, DLRM, transformer). State the latency/accuracy trade-off explicitly.
-
7
Serving & scale
Candidate generation vs ranking, ANN retrieval, caching, batching, and a heavy-then-light multi-stage funnel. Cover cold start and fallbacks.
-
8
Monitoring & iteration
Drift detection, guardrail metrics, online evaluation, retraining cadence, and failure modes. End by listing what you'd build next.
Worked case studies
Frame it
- Two-stage funnel: candidate generation (retrieve ~1000 from millions) then ranking (score the 1000).
- Frame ranking as predicting P(engagement) — watch, click, purchase.
Data & labels
- Implicit feedback: clicks, watch-time, purchases as positives; impressions-without-action as negatives.
- Beware position bias — top items get more clicks regardless of relevance.
Model
- Retrieval: two-tower (user tower + item tower) with embeddings + approximate nearest neighbor (ANN / FAISS).
- Ranking: gradient-boosted trees or a DLRM/DCN deep model on rich features.
Serving & scale
- Precompute item embeddings offline; compute user embedding online.
- ANN index for sub-50ms retrieval; cache popular results; fallback to trending for cold-start users.
Metrics
- Offline: recall@k, NDCG. Online: CTR, watch-time, retention via A/B test.
- Add diversity / freshness guardrails so the feed doesn't collapse to one genre.
Pitfalls
- Feedback loops reinforce popular items; inject exploration (epsilon-greedy / bandits).
- Filter-bubble and stale-embedding drift; retrain regularly.
Frame it
- Multi-objective ranking: predict P(like), P(comment), P(share), P(hide) and combine with a weighted value model.
- Score = w1·like + w2·comment + w3·share − w4·hide.
Features
- User-author affinity, post recency, content type, historical engagement, social graph signals.
- Real-time features (last session activity) via a feature store.
Model
- Multi-task deep network with shared bottom + per-objective heads (MMoE).
- Calibrated probabilities so the weighted combination is meaningful.
Metrics
- Online: time-spent, meaningful interactions, day-N retention.
- Guardrails: integrity (downrank misinformation), reported-content rate.
Pitfalls
- Optimizing pure engagement can amplify clickbait/outrage — needs integrity objectives.
- Cold-start for new users and new posts.
Frame it
- Binary classification: P(click | user, ad, context). Expected value = bid × pCTR drives ranking.
- Massive sparse categorical features (billions of feature crosses).
Model
- Logistic regression with hashing as a baseline; then Wide & Deep, DeepFM, or DCN for feature interactions.
- Embeddings for high-cardinality IDs (user, ad, advertiser).
Data
- Severe class imbalance (clicks are rare ~1-5%); use negative downsampling + calibration.
- Delayed feedback — a click may arrive minutes later.
Metrics
- Offline: AUC, log-loss, calibration (predicted vs actual CTR).
- Online: revenue, click yield, advertiser ROI.
Pitfalls
- Calibration matters more than ranking accuracy because pCTR feeds a money auction.
- Position bias; train with position as a feature, serve at fixed position.
Frame it
- Binary classification with extreme imbalance (fraud << 0.1%) and a hard real-time latency budget.
- Cost-sensitive: a missed fraud costs far more than a false alarm.
Data & labels
- Labels are delayed and noisy (chargebacks arrive weeks later).
- Velocity / aggregate features: # transactions last hour, distance from last location, amount vs user baseline.
Model
- Gradient-boosted trees for tabular features; graph features to catch fraud rings.
- Unsupervised anomaly detection (isolation forest / autoencoder) to catch novel patterns.
Serving
- Real-time scoring < 100ms inline with the transaction; rules engine for hard blocks + ML score for the gray zone.
- Human review queue for borderline cases generates new labels.
Metrics
- Precision/recall at a fixed alert budget, dollars saved, false-positive rate (customer friction).
- Use PR-AUC, not ROC-AUC, under heavy imbalance.
Pitfalls
- Adversaries adapt — concept drift is constant; retrain frequently.
- Feedback loop: blocked transactions never get a fraud label.
Frame it
- Retrieval then learning-to-rank. Query understanding + candidate retrieval + re-ranking.
- Frame re-ranking as predicting relevance (pointwise / pairwise / listwise).
Retrieval
- Lexical (BM25 / inverted index) + semantic (embedding + ANN) — hybrid retrieval beats either alone.
- Query expansion and spell correction up front.
Model
- LambdaMART (GBT) is still a strong LTR baseline; cross-encoder transformer for top-k re-rank.
- Features: text match, popularity, freshness, personalization, click history.
Metrics
- Offline: NDCG, MRR, recall@k from human-judged or click-derived labels.
- Online: click-through, successful-session rate, time-to-success.
Pitfalls
- Click labels are biased toward what was shown and ranked high.
- Cross-encoders are accurate but slow — only re-rank a small top-k.
Frame it
- Not a from-scratch model — orchestrate retrieval + a foundation LLM with grounding.
- Goal: accurate, cited answers with low hallucination.
Pipeline
- Ingest → chunk → embed → vector DB. At query time: embed query → retrieve top-k chunks → stuff into prompt → generate.
- Add re-ranking of retrieved chunks and a citation step.
Key choices
- Chunk size/overlap, embedding model, hybrid (keyword+vector) retrieval, context-window budget.
- Guardrails: refuse when retrieval confidence is low; cite sources.
Metrics
- Retrieval: recall@k, MRR. Generation: faithfulness/groundedness, answer relevance (LLM-as-judge + human eval).
- Online: thumbs-up rate, deflection rate, escalation rate.
Pitfalls
- Hallucination when retrieval misses; stale index; prompt injection from documents.
- Latency and cost of long contexts — cache embeddings and frequent answers.
Frame it
- Binary classification over a prediction window; define churn precisely (no activity 30d? cancellation event?).
- Output a risk score that triggers a retention intervention.
Features
- Engagement trend (declining usage is the strongest signal), tenure, support tickets, billing events, RFM.
- Build features as of a cutoff date to avoid label leakage.
Model
- Gradient-boosted trees on tabular features; survival analysis if you care about time-to-churn.
- Calibrate scores so retention budget targets the right users.
Metrics
- Offline: PR-AUC, lift@decile. Business: retained revenue from interventions (measured via holdout).
- Uplift modeling: target users the intervention actually changes, not just high-risk users.
Pitfalls
- Leakage from features recorded after the churn decision.
- Acting on the score changes future data (intervention effect).
Frame it
- Regression on travel time; often predict a distribution (quantiles), not a point, to set expectations.
- Decompose: route time + pickup/handoff time + queueing.
Features
- Distance, historical segment speeds, time-of-day, weather, traffic, driver/restaurant state.
- Real-time signals (current congestion) via streaming features.
Model
- GBT baseline; graph neural nets / road-segment models for routing-aware estimates.
- Quantile loss for under-promise/over-deliver behavior.
Metrics
- MAE / quantile loss; % of arrivals within the promised window.
- Asymmetric cost: being late is worse than being early.
Pitfalls
- Self-fulfilling: shown ETA changes user/driver behavior.
- Long-tail events (accidents, surges) dominate user dissatisfaction.
Frame it
- Multi-label classification across policy categories; very high recall on severe categories.
- Multi-stage: cheap filter → ML model → human review for the uncertain band.
Model
- Fine-tuned vision/text transformers; multimodal for memes (image+text).
- Hash-matching (PhotoDNA-style) for known-bad content before the model.
Data
- Severe imbalance and shifting adversarial content; active learning on borderline cases.
- Human-labeled golden set per policy with clear guidelines.
Metrics
- Recall on harmful content (don't miss), precision (don't over-remove), appeal-overturn rate.
- Per-category thresholds tuned to harm severity.
Pitfalls
- Adversarial evasion (obfuscated text, cropped images); context matters (satire vs hate).
- Reviewer well-being and consistency.
Frame it
- Time-series regression at scale (thousands of series); predict horizon with uncertainty.
- Hierarchical: forecasts should reconcile across product/store/region levels.
Features
- Lags, rolling stats, seasonality (weekly/yearly), holidays, promotions, price, weather.
- Avoid leakage — only use information available at forecast time.
Model
- Baselines: seasonal naive, ETS/ARIMA. Scale: gradient-boosted trees (LightGBM) on engineered features, or global deep models (DeepAR / Temporal Fusion Transformer).
- Quantile forecasts for safety stock.
Metrics
- MAPE / WAPE / pinball loss; bias (consistent over/under-forecasting hurts inventory).
- Backtest with rolling-origin evaluation, never random splits.
Pitfalls
- Intermittent/zero-heavy demand; cold-start for new products.
- Promotions and stockouts distort the historical signal.