The ML Interview Book¶
From first principles to the frontier, for staff-level ML interviews.
Full table of contents: every part and chapter on one page.
This book is organised as a compendium rather than a syllabus. Every chapter connects one topic downward to its mathematics and upward to the modern foundation-model stack, until you can move freely between
\[
\text{equation} \rightarrow \text{implementation} \rightarrow \text{architecture} \rightarrow \text{training dynamics} \rightarrow \text{distributed system} \rightarrow \text{evaluation} \rightarrow \text{production case study.}
\]
Three standards are applied to every topic:
| Standard | The bar |
|---|---|
| Math | Derive and explain the important equations without hand-waving. |
| Code | Implement the core algorithm without calling a library implementation, with every tensor shape written down. |
| Systems | Explain when it works, when it fails, its compute and memory characteristics, and how a real company trained, deployed, and evaluated it. |
How the book is organised¶
flowchart TB
A[I. Math] --> B[II. Classical ML] --> C[III. Neural nets]
C --> D[IV. Vision] --> H[VIII. Multimodal]
C --> E[V. Sequences & Transformers] --> F[VI. LLM training] --> G[VII. Post-training]
H --> K[XI. Perception & autonomy]
C --> I[IX. Generative] --> K
C --> J[X. Self-/semi-/weak supervision] --> K
L[XII. Reinforcement learning] --> G
L --> K
M[XIII. Retrieval, eval, reliability] --> Q[XVII. ML system design]
N[XIV. Systems] --> Q
O[XV. Interpretability & safety] --> Q
P[XVI. Coding canon] --> Q --> R[XVIII. Company deep dives]- Parts I–III are the substrate: linear algebra, probability, optimization, and a neural network built from scratch in NumPy, including an autograd engine.
- Parts IV–XI are the five pillars for a perception + multimodal + LLM engineer: vision, Transformers, LLM training, post-training / RL, and autonomous perception.
- Parts XII–XV are what separates a senior engineer from a staff engineer: RL from the MDP up, retrieval and evaluation, distributed training and inference systems, interpretability and failure modes.
- Part XVI is the coding canon: 60 implementations you should be able to write from memory, with NumPy ↔ PyTorch fluency drills.
- Part XVII is ML system design, structured the way a staff-level interview is graded: requirements, trade-offs, a committed decision, and the reasoning behind it.
- Part XVIII personalises that to the companies you are likely to interview with, using their actual business problems and their own published engineering write-ups.
Every chapter ends with interview questions (with model answers that show reasoning,
not recall) and exercises with solutions. All code lives in src/mlbook/ and is tested
in tests/.
Reading modes¶
- Cover to cover. follow the study plan.
- The night before (read each chapter's TL;DR) the interview card.
- Coding rounds. Part XVI and the tests in
tests/. - System design rounds. Part XVII, then the company deep dive.