Data Science & ML Interview Cheat Sheet & Architectures
Master Machine Learning System Design, LLM RAG pipelines, MLOps, Loss Functions, and A/B Testing statistics for senior and staff Data Science & ML Engineering loops.
The 7-Step ML System Design Framework
The standard blueprint evaluated in senior and staff ML engineering rounds.
- 01
Problem Formulation
Clarify business objective, convert to ML task (Classification, Ranking, Regression), and define offline vs online metrics.
- 02
Data Pipeline & Features
Feature store schema, point-in-time joins to prevent data leakage, categorical embeddings, and real-time streaming features.
- 03
Candidate Retrieval (Stage 1)
Two-tower embedding models (User & Item towers) retrieving top-K items via Approximate Nearest Neighbor (ANN) vector search.
- 04
Heavy Ranking (Stage 2)
Deep neural network with cross-features, multi-gate mixture-of-experts (MMoE), and calibrated probability outputs.
- 05
Re-ranking & Diversity (Stage 3)
Deduplication, business rules, freshness boosts, and exploration-exploitation (Multi-Armed Bandits / LinUCB).
- 06
Serving & Latency SLAs
Model quantization (INT8), ONNX Runtime/TensorRT, model cache layers, and sub-50ms p99 SLA guarantees.
- 07
Monitoring & Feedback Loop
Concept drift detection (PSI / KL divergence), training-serving skew alerts, and automated continuous re-training triggers.
Essential Statistics & A/B Testing Formulas
Core concepts for metric evaluation, hypothesis testing, and causal inference.
| Concept | Definition / Rule | Interview Context |
|---|---|---|
| Sample Size & Power | N = 16 × σ² ÷ MDE² (for α=0.05, β=0.8) | Determine experiment runtime to detect Minimum Detectable Effect |
| Novelty Effect | Initial spike from curiosity that decays over time | Run experiments for full business cycles (minimum 2 weeks) |
| Network Interference | Treatment spills over to control (SUTVA violation) | Cluster-based or switchback experimentation (e.g. Uber/Airbnb) |
Frequently Asked Questions on Data Science & ML Interviews
Insights on modeling depth, MLOps, and architecture interview preparation.
What are the major interview rounds in a Data Science & ML Engineering loop?
DS/ML loops evaluate 4 main areas: ML System Design (Recommenders, Search Ranking, LLM RAG), Applied Machine Learning & Modeling (Loss functions, Bias-Variance trade-off, Regularization), Coding & SQL/Data Wrangling, and Experimentation & Statistics (A/B testing, Power Analysis, Causal Inference).
What is the standard architecture for an ML recommendation system?
A production recommendation system uses a two-stage architecture: Candidate Retrieval / Candidate Generation (Vector search with approximate nearest neighbors using HNSW/Faiss to narrow millions of items to ~1,000) followed by Heavy Ranking (Multi-task deep neural network predicting CTR, watch time, and conversion).
How does ClawPad assist ML and Data Science candidates in interviews?
ClawPad renders real-time SVG diagrams of feature stores, training/serving skew prevention pipelines, two-tower embedding models, and vector database architectures as the answer streams.