Free ML Interview Prep

All 10 tracks from deep-ml.com/interview-prep, rebuilt with links to free, canonical resources on the open web. No premium paywall.

183 questions · 286 citations · 7 curated tracks · 100% free

Anthropic

AI-First · MLE, Research, Data Scientist

This track prepares you for Anthropic's Machine Learning Engineer loop, which weighs clean modular Python, concurrency, and from-scratch ML coding alongside heavy ML-systems design (inference/serving, distributed training, eval and safety pipelines) and ML/RL theory with a project deep-dive. Content is prioritized by real interview frequency: high-frequency items are core, mediums recommended, lows stretch.

Interview process

  1. Recruiter Screen & Hiring-Manager Technical Screen~30 min + 45-60 min

    A non-trivial recruiter call (motivation, background, 'why Anthropic', mission alignment) and a hiring-manager technical deep-dive into your past ML projects and decisions. The recruiter often names the coding prompt family in advance.

  2. CodeSignal OA + Live Coding / ML-Ops90 min OA (4 levels) + 1-2 x 55-60 min

    A four-level CodeSignal assessment (one 'code that evolves' problem, ~600-pt scale) plus live rounds: practical build-from-scratch and concurrency in Python, an MLE Python/ML-ops task (data cleaning or improving code from an internal tool/agent loop), and ML-coding-from-scratch (attention, BPE, sampling) and debugging. Clean, modular, runnable code matters more than algorithmic tricks. AI tools banned; internet allowed.

  3. ML / System Design55-60 min (shared Google Doc)

    The heart of the MLE loop: inference/serving APIs (batching, KV cache, GPU memory, streaming), distributed-training systems (data/tensor/pipeline parallelism, FSDP, fault tolerance), eval pipelines, data pipelines, retrieval/search at scale, and safety-classifier pipelines.

  4. ML Theory / RL Fundamentals & Project Deep-Dive55-60 min

    ML/RL theory (scaling laws, transformer internals, RL fundamentals incl. on/off-policy and advantage estimation) and a technical project deep-dive on an ML system you built end-to-end (decisions, tradeoffs, failures). Lighter on novel-research presentation than the RS track.

Free resources

OpenAI

AI-First · Data Scientist, Research, MLE

OpenAI hires Data Scientists who pair rigorous experimentation and product analytics with fluency in AI products and AI tooling. The loop blends a product-analytics take-home, live SQL/Python, experiment/product-sense case studies, and PhD-level statistics plus LLM-evaluation depth. This track maps each real question to hands-on Deep-ML practice or focused concept prep.

Interview process

  1. Recruiter & Hiring-Manager Screen~30-45 min + ~30 min

    A non-technical recruiter call (background, motivations, 'why OpenAI', and - distinctively - your experience working with the output of AI coding tools; enterprise DS experience for some teams), then sometimes a hiring-manager chat on your interest in AI/OpenAI and a light project discussion.

  2. Take-Home Data Challenge & Technical Review48-hr take-home + ~60 min review

    A 48-hour product-analytics take-home (interpret an A/B experiment, do segmentation and EDA, communicate via slides) where using AI tools like ChatGPT is explicitly encouraged. Then a one-hour review: ~30 min presenting your deck, ~30 min on live SQL plus reading AI-generated/experiment-setup code and fixing what it got wrong.

  3. Case Studies, Experimentation & Product Sense2 case studies + PM case (final loop)

    Two data-science case studies (given feature metrics: analyze, recommend launch/no-launch, find the p-value, design a multifactorial experiment) plus a project-management case (a shift in user behavior; lay out your approach under time/resource constraints and discuss collaboration and stakeholder communication).

  4. Statistics Q&A & ML/LLM-Evaluation Depth45-60 min (final loop)

    A statistics/experimentation Q&A round with academic/'PhD-level' questions applied to real problems (p-values, power, causal inference), plus AI-product-flavored depth: measuring LLM output quality, evaluating/comparing models, analyzing bias, and building eval frameworks.

Free resources

Google DeepMind

AI-First · RS, RE

Google DeepMind's LLM Researcher (RS/RE) interviews probe deep transformer internals, pretraining/scaling-law reasoning, alignment (RLHF/DPO), and large-scale training+serving systems, alongside a no-AI ML coding bar and a math/probability quiz. This track maps each real question to hands-on Deep-ML practice plus targeted explain-and-design concepts.

Interview process

  1. Recruiter Screen30 min

    Background, motivation, and a couple of quick ML fundamentals.

  2. ML Coding Onsite60 min

    Two gated rounds of LeetCode medium-hard plus from-scratch ML implementations. Code must actually run, and AI assistants are banned.

  3. ML System Design60 min

    Open-ended design of an end-to-end ML system, justified from first principles (data, architecture, training, serving, evaluation).

  4. Research Discussion / Behavioral45 min

    Paper deep-dive and project defense, plus deep ML/math theory and derivations (the technical 'quiz' depth is probed here in this format).

Free resources

Meta

Big Tech · Data Scientist, MLE

Meta's Data Scientist (Analytics) loop rewards relational SQL fluency, sharp product intuition, and rigorous experiment design over heavy ML. This track drills the exact SQL patterns (streaks, retention, window functions, ratios), the ambiguous product-sense cases, and the applied-statistics/A-B-testing reasoning Meta interviewers push on.

Free resources

NVIDIA

Big Tech · AI Engineer, GPU perf, MLLM systems

NVIDIA AI Engineer loops pair rigorous fundamentals (gradient-descent theory, normalization, optimizers, transformers) with hands-on from-scratch ML implementations and, depending on team, either GPU/CUDA performance design or LLM/MLLM system design. Expect fast LeetCode-medium coding plus a signature 'implement this ML primitive live' round and deep 'blast' questions on optimizers and normalization.

Free resources

Jane Street

Finance · MLE, MLR

Jane Street hires ML Engineers/Researchers who can defend every modeling choice on noisy, non-stationary financial data, implement ML and data-structure code cleanly and quickly under layered follow-ups, and reason sharply about probability and expected value. This track pairs from-scratch implementations and streaming/statistics drills with deep modeling-theory and systems-design discussion.

Free resources

Machine Learning Engineer (General)

Other · MLE

A general MLE track: ML theory, classic algos, DL/LLM basics, prob/stats, from-scratch coding, ML system design, and production MLOps.

Free resources

ML System Design (General)

Other · Free on deep-ml

End-to-end production ML: recommendation/search/ads/trust-and-safety/LLM systems, defending metric, data, and scaling trade-offs.

Foundations & Framing

  1. Concept Walk me through your framework for approaching any ML system design problem.
  2. Concept How would you translate this vague business goal into a well-defined ML problem?
  3. Concept What clarifying questions do you ask before you start designing?
  4. Concept Would you even use ML for this, or would a rule-based system suffice?
  5. Concept How do you choose success metrics for this system?
  6. Concept Estimate the scale: QPS, data volume, storage, and the latency budget.
  7. Concept How would a staff-level answer to this problem differ from a mid-level one?
  8. Concept Why start with a simple baseline, and why logistic regression before GBDT before deep nets?
  9. Concept When would you pick GBDT over a deep model (and vice versa)?
  10. Concept Why use candidate generation + ranking instead of one model?

Data, Metrics & Training Fundamentals

  1. Concept Engineer the features (user / item / context / cross) for this model.
  2. Concept How would you build the labeling / training-data-generation pipeline?
  3. Concept Implement online + offline feature computation with consistency (and <10ms online serving).
  4. Concept Design the retraining pipeline (reproducibility, failure recovery, hyperparameter tuning).
  5. Concept Why might offline metrics improve while online metrics don't?
  6. Concept When do you use NDCG vs MAP vs MRR? AUC vs PR-AUC?
  7. Concept Why normalized cross-entropy for CTR, and what is calibration?
  8. Concept How do you generate negatives when you only observe positive signals?
  9. Concept What are the trade-offs of implicit vs explicit feedback labels?
  10. Concept Why should you never use random splits on time-series interaction data?
  11. Concept How do you handle severe class imbalance (e.g., fraud at <0.1%)?
  12. Concept How does a two-tower retrieval model work, and when do you use it?

Core Design: Recsys, Search & Ads

  1. Concept Design YouTube / Netflix video recommendations.
  2. Concept Design the Facebook / Instagram news feed ranking system.
  3. Concept Design Instagram Reels / TikTok short-video recommendation.
  4. Concept Design 'People You May Know' / a friend recommendation system.
  5. Concept Design Amazon / e-commerce product recommendation.
  6. Concept Design Airbnb listing ranking / similar-listing recommendations.
  7. Concept Design an event recommendation system (e.g., Eventbrite).
  8. Concept Design a related-pins / similar-items system (Pinterest).
  9. Concept Design a search ranking system.
  10. Concept Design product search ranking for an e-commerce platform.
  11. Concept Design a visual / image search system.
  12. Concept Design autocomplete / typeahead / query suggestion.
  13. Concept Design a semantic search system.
  14. Concept Design an ad click-through-rate (CTR) prediction system.
  15. Concept Design an ad ranking and auction system.
  16. Concept Design a conversion (pCVR) prediction system.
  17. Concept How do you handle new users and new items (cold start)?
  18. Concept How do position bias and popularity bias arise, and how do you correct them?
  19. Concept How do feedback loops / filter bubbles form, and how do you break them?
  20. Concept When would you use bandits / exploration in a recommender?

Trust, CV/NLP, Marketplace & Serving Ops

  1. Concept Design a harmful / abusive content detection system.
  2. Concept Design a spam / abuse detection system.
  3. Concept Design a payments fraud detection system (Stripe-like).
  4. Concept Design a bot / fake-account detection system.
  5. Concept Design a general content moderation pipeline.
  6. Concept Design an image classification service.
  7. Concept Design an object detection system.
  8. Concept Design a Google Street View blurring system (faces / license plates).
  9. Concept Design a machine translation system.
  10. Concept Design a speech recognition / voice assistant system.
  11. Concept Design an ETA prediction system (Uber / DoorDash).
  12. Concept Design a dynamic / surge pricing system.
  13. Concept Design a demand forecasting system.
  14. Concept Design a delivery / dispatch or rider-matching system.
  15. Concept Design a churn prediction system.
  16. Concept Design a price / value estimation model (Zillow home price, pay-range).
  17. Concept Data drift vs concept drift - how do you detect each, and when do you retrain?
  18. Concept What is training-serving skew and how do you prevent it?
  19. Concept Compare shadow deployment, canary, A/B rollout, and rollback.
  20. Concept When do you quantize, prune, or distill a model?
  21. Concept Decompose a sub-100ms serving latency budget.
  22. Concept How does your design change at 10x / 100x / 1000x scale?
  23. Concept How do you address fairness, bias, and privacy in this system?

LLM/GenAI & ML Infra

  1. Concept Design a chatbot / AI customer-support assistant.
  2. Concept Design an LLM evaluation pipeline.
  3. Concept Design an agentic system with tool use.
  4. Concept Design a hallucination detection / guardrails system for LLM outputs.
  5. Concept Design a code-completion system (Copilot).
  6. Concept Design a semantic-caching layer for LLM APIs.
  7. Concept Design a feature store.
  8. Concept Design training for a model that doesn't fit on a single GPU.
  9. Concept Design a model-serving platform.
  10. Concept Design an experimentation / A-B testing platform.
  11. Concept Design an ML monitoring / observability system.
  12. Concept Design a vector database.

Inference Engineer (General)

Other · Free on deep-ml

LLM inference: prefill/decode economics, KV-cache math, attention kernels, quantization numerics, parallelism, and end-to-end serving system design.

Inference Fundamentals & Numerics

  1. Concept Explain prefill vs decode. Why is prefill compute-bound and decode memory-bandwidth-bound?
  2. Concept Walk through the roofline model for LLM inference. How do you determine whether a workload is memory- or compute-bound?
  3. Concept Define TTFT, ITL/TPOT, and throughput. How do they trade off, and which do you optimize for interactive vs batch workloads?
  4. Concept You serve a large model and throughput barely improves when you double the batch size. Why, and what would you do about it?
  5. Concept Compare FP32, TF32, FP16, BF16, and FP8 - when to use each, and why numerical precision matters for inference correctness.
  6. Concept Derive backprop for softmax + cross-entropy cleanly, and explain the transformer attention computation end to end.

KV Cache & Attention Internals

  1. Concept What is the KV cache, and why does it turn decode from quadratic into linear work per token?
  2. Concept Explain PagedAttention. What operating-system concept does it borrow, and which problem does it solve?
  3. Concept Compare MHA, MQA, GQA, and MLA - and quantify their effect on KV-cache size.
  4. Concept What is prefix / radix caching, and when does cross-request KV reuse actually pay off?
  5. Concept How would you quantize the KV cache, and what accuracy and acceptance-rate risks does that introduce?
  6. Concept Explain sliding-window / streaming attention and KV eviction policies for long context.
  7. Concept Explain FlashAttention: the IO-aware argument, tiling, online softmax, and recomputation. What changed in v2/v3?

Batching, Decoding & Quantization

  1. Concept Contrast static, dynamic, and continuous / in-flight batching. Why does continuous batching help so much for LLM decode?
  2. Concept Design an inference batching system for a single GPU that handles up to N concurrent inputs while users wait synchronously, maximizing utilization.
  3. Concept What is chunked prefill, and how does it protect decode latency from long prompts?
  4. Concept Explain prefill-decode disaggregation. Why might you run prefill and decode on separate hardware pools?
  5. Concept How do you handle request scheduling, preemption, fairness/priority, and admission control under bursty load?
  6. Concept Explain the difference between weight-only and weight+activation quantization, and why activation outliers make the latter hard.
  7. Concept Compare GPTQ, AWQ, and SmoothQuant. When would you pick each, and which are weight-only?
  8. Concept Compare FP8 (E4M3 vs E5M2), INT8, INT4, and FP4 - dynamic range, accuracy risk, and hardware support.
  9. Concept How do you evaluate accuracy-vs-speed after quantizing? What regressions would worry you, and how would you catch them before production?
  10. Concept Explain speculative decoding. Why does it only help when the draft model is cheap enough, and what is the acceptance rate?
  11. Concept Explain structured / constrained decoding (JSON, grammars) and its latency cost.

Parallelism & Serving Systems

  1. Concept Diagnose whether a kernel is memory-bound or compute-bound and describe how you'd profile it (Nsight, torch profiler).
  2. Concept Explain tensor parallelism vs pipeline parallelism. When do you use each for inference rather than training?
  3. Concept Serve a model that doesn't fit on one GPU across N GPUs - how do you shard it, and what are the communication costs?
  4. Concept Expert parallelism / MoE serving - what changes about routing, memory, and load balance?
  5. Concept Design an LLM serving platform / inference gateway for X QPS at a Y p99 target - request routing, admission control, batching, KV management, observability, failover.
  6. Concept Design KV-cache-aware routing and load balancing across replicas. Why does naive round-robin hurt?
  7. Concept How do you handle cold starts and model loading? Discuss scale-to-zero tradeoffs and GPU memory fragmentation in multi-model serving.
  8. Concept What do you track and alert on in production inference (p50/p95/p99 TTFT and ITL, GPU utilization, queue depth, cost per token)?
  9. Concept Estimate cost per token / GPU-hour and do capacity planning for a launch. Then: cut cost per token by 50% - what levers do you pull, in what order?
  10. Concept Design safe model rollout: canary and shadow deploys, versioning, and rollback for an inference fleet.
  11. Concept Compare vLLM, SGLang, TensorRT-LLM, and TGI. What does each do best, and how would you choose for a given workload?
  12. Concept What do torch.compile / TorchInductor, ONNX export, and CUDA graphs optimize, and what are the common pitfalls?
  13. Concept Deploy a vision/CNN model to production: batch vs real-time, ONNX conversion pitfalls, and edge/CPU/on-device constraints.

Debugging, Coding & Behavioral

  1. Concept Throughput dropped 30% after a model or config update - debug it.
  2. Concept GPU utilization is sitting at 40% - why, and how do you raise it?
  3. Concept p99 latency spikes intermittently under steady load - diagnose it.
  4. Concept You're hitting OOM under load - find and fix it (KV growth, fragmentation, batch sizing, long contexts).
  5. Concept Tell me about an incident you were on call for - how did you debug it, and what did you change to prevent recurrence?
  6. Concept Describe a hard tradeoff you made between latency, cost, and accuracy - and how you decided. Also: a time you worked with a research team to productionize a model.

Data Science (General)

Other · Free on deep-ml

Data Scientist interviews blend statistical and probabilistic reasoning, SQL/pandas data manipulation, ML theory, and product-sense A/B testing.

Foundations: Statistics & Probability

  1. Concept What is a p-value? Explain it precisely (and why it is NOT the probability that the null hypothesis is true).
  2. Concept What is a confidence interval? How do you interpret a 95% CI (and what does it NOT mean)?
  3. Concept Explain Type I vs Type II error and statistical power. What is worse depends on context - discuss.
  4. Concept Explain the Central Limit Theorem and why it matters in practice.
  5. Concept What is the Law of Large Numbers and how is it used?
  6. Concept Compare Bernoulli, Binomial, Poisson, Normal, Uniform, and Geometric distributions; when does each arise?
  7. Concept For a right-skewed distribution, which is larger, the mean or the median? Why?
  8. Concept What is the difference between probability and likelihood?
  9. Concept Difference between the frequentist and Bayesian approaches to inference.
  10. Concept State and apply Bayes' theorem (e.g., a disease test with given sensitivity/specificity, or fair vs unfair coin).
  11. Concept What is selection/sampling bias, and how would you detect and correct for it?
  12. Concept Correlation vs causation - how would you establish causation when you cannot run an experiment?
  13. Concept What is the expected value of a single dice roll? When should you choose to re-roll a die to maximize expected value?
  14. Concept Two children problem: given that one child is a boy (or girl), what is the probability both are the same sex?
  15. Concept Monty Hall problem: should you switch or stay, and why?
  16. Concept There is a fair coin and an unfair coin (both sides tails). You pick one, flip 5 times, and get 5 tails. What is the probability you picked the unfair coin?
  17. Concept How do you generate fair 50/50 odds from a biased coin (von Neumann trick)?
  18. Concept What is the expected number of coin flips to get two consecutive heads?
  19. Concept How many cards do you expect to draw from a shuffled deck before you see the first ace?
  20. Concept What is the difference between a z-test and a t-test, and when do you use each?

SQL & Data Manipulation

  1. Concept Explain the difference between WHERE and HAVING, and how GROUP BY aggregations work.
  2. Concept Explain INNER vs LEFT JOIN and UNION vs UNION ALL; how would you explain a JOIN simply?

Python & Pandas

  1. Concept Explain .loc vs .iloc, and the difference between a NumPy array and a Python list.

Machine Learning Theory

  1. Concept Difference between supervised and unsupervised learning, with examples of each.
  2. Concept Explain the bias-variance tradeoff.
  3. Concept What is overfitting, and how do you detect and prevent it?
  4. Concept When and why do you scale features (normalization vs standardization)? Which models need it?
  5. Concept L1 (Lasso) vs L2 (Ridge) regularization - the difference, and when to use each.
  6. Concept Explain cross-validation, including k-fold and stratified k-fold.
  7. Concept Define precision, recall, and F1, and explain the precision-recall tradeoff.
  8. Concept ROC-AUC vs PR-AUC, and when is accuracy a misleading metric?
  9. Concept How do you handle imbalanced data (resampling, class weights, threshold moving, choice of metric)?
  10. Concept How do you handle missing data (deletion vs imputation; MCAR/MAR/MNAR)?
  11. Concept Describe feature-selection methods (filter, wrapper, embedded) and give examples of feature engineering.
  12. Concept What are the assumptions of linear regression, and how do you check them?
  13. Concept Explain logistic regression and its loss function (log loss).
  14. Concept How does a decision tree work, and how do random forests reduce overfitting?
  15. Concept Explain the difference between bagging and boosting (and how gradient boosting / XGBoost works).
  16. Concept Explain k-means clustering and how you would evaluate cluster quality (e.g., silhouette score, elbow).
  17. Concept Explain PCA and when you would NOT use it.
  18. Concept Explain KNN and how the choice of k affects the bias-variance tradeoff.
  19. Concept What is data leakage, how does it happen, and how do you prevent it?
  20. Concept How would you explain a model and its predictions to a non-technical stakeholder?

Experimentation, Product Sense & Behavioral

  1. Concept How would you measure the success of a product or feature (define the success metrics)?
  2. Concept What are guardrail metrics and why do you track them?
  3. Concept How would you design an A/B test for a new feature or homepage layout?
  4. Concept How do you determine the sample size and duration for an experiment (effect size, significance, power)?
  5. Concept Explain the multiple-testing problem and corrections (e.g., Bonferroni); what is p-hacking?
  6. Concept What is the novelty effect, and what is the problem with peeking (a PM wants to check results on day 3)?
  7. Concept What do you do if an A/B test result is not statistically significant?
  8. Concept A key metric dropped (e.g., DAU dropped 5%). How do you investigate?
  9. Concept Two metrics conflict (engagement up but revenue flat, or one region up and another down). What do you do?
  10. Concept Diagnose a discrepancy across platforms or segments (e.g., iOS vs Android adoption differs).
  11. Concept How do you handle network effects / interference in a two-sided marketplace or social network experiment?
  12. Concept When should you NOT run an A/B test, and what alternatives would you use (e.g., quasi-experiments, causal inference)?
  13. Concept How would you quantify the influence of a user (e.g., a Twitter user)?
  14. Concept Walk me through a past data science project end-to-end, and explain why you chose your approach/model.
  15. Concept Tell me about a time you explained a complex technical result to a non-technical stakeholder.
  16. Concept Describe a time you dealt with an ambiguous or poorly-defined problem.
  17. Concept Tell me about a project that failed or did not go as planned, and what you learned.
  18. Concept How do you prioritize competing requests or multiple projects?
  19. Concept Tell me about a time you influenced a decision with data, or defended your findings to skeptical stakeholders.
  20. Concept Describe a time you worked with a cross-functional team (PMs, engineers).