Build status

Markdown in content/ builds to HTML in _site/. Multi-course aware: adding a course needs a directory, a course.yml, and Markdown files — no code change.

A shared editorial reading layout with per-topic accent colours, Georgia headings, DM Sans text, Fira Code, and light, dark, or system-driven appearance.

KaTeX for inline and display math; Mermaid diagrams with pan, zoom, and full-size view; syntax-labelled code blocks; scrollable tables.

Pushes to main build and publish to GitHub Pages automatically. Pull requests run the same build without deploying.

Every deploy is gated on a check suite asserting that each source file produced a page, internal links resolve, diagram containers exist, and math markup survived the pipeline. Browser rendering and semantic correctness require separate checks.

mlinterviewnotes.com, served over TLS.

An earlier review covered 67 Transformer diagrams. The later curriculum audit identified additional mathematical and implementation qualifications; that historical pass is not a current guarantee of complete diagram or scientific correctness.

Pages can embed raw HTML blocks and pull in their own scripts, so D3-driven explanations live alongside ordinary Markdown.

Full-text client-side search across courses, notes, and labs, with section filters and saved-page browsing. No backend needed.

Adjustable text size, focus mode, saved pages, reading progress, resume reading, and code copying. Preferences stay in the browser.

Each of the 24 ML/DL topic pages has independent CPU examples, with downloadable Python generated from the displayed code. A separate test suite checks coverage, assertions, execution and source parity; GPU and distributed fragments are outside that suite's scope.

From "why did we abandon RNNs?" through to the configuration choices in production 2026 models. Includes 67 diagrams, worked numeric examples, self-check questions, and a runnable reference decoder.

How LLM serving systems actually work, read out of the vLLM and SGLang source: continuous batching, PagedAttention and RadixAttention, quantization, parallelism, speculative decoding, and honest benchmarking. 91 pages with 304 diagrams and hands-on labs. Ships as prebuilt HTML re-themed to the site palette — see the note below.

PyTorch 2.8 from tensor strides and autograd internals through torch.compile, distributed training, custom kernels, export and quantization, ending with a GPT-style model built, trained, evaluated, fine-tuned and exported on a CPU. Every module carries executable checks, and the labs are tested in CI.

From "why not just prompt it?" through supervised fine-tuning, LoRA and QLoRA, preference optimisation, RL with verifiable rewards and distillation, to agents, multimodal and long-context tuning, merging, safety, and serving adapters at a known cost. Worked numeric examples, self-check questions, an interview bank, and executable CPU checks.

Linear algebra (with five interactive visualizations), calculus and matrix calculus, probability, statistics and inference, optimization, information theory, discrete mathematics, and numerical computing. Worked examples, diagrams, failure modes, and self-check questions are published. The audit's corrections and further depth work remain.

NumPy, pandas, scikit-learn, PyTorch, TensorFlow and Keras, the boosting libraries, visualization, the Hugging Face stack, and MLOps and serving. Each page teaches the library's underlying model — memory layout, index alignment, the estimator contract, tracing, the fit/transform boundary — rather than listing API calls.

Problem framing, linear models, trees and ensembles, SVMs and kernels, probabilistic and instance-based models, unsupervised learning, bias-variance and generalization, model evaluation, feature engineering, hyperparameter tuning, imbalanced data and practical pitfalls, and ethics and fairness. Every chapter has runnable CPU examples, worked calculations and solved self-checks. Linear/logistic regression, decision trees, random forests, KNN, KMeans, AdaBoost, XGBoost, CatBoost and hierarchical clustering have explicit coverage.

Neural networks and backpropagation, activations and initialization, optimization and training, regularization and normalization, CNNs, RNNs and sequence models, attention and transformers, transfer learning and PEFT, generative models, self-supervised learning, and deep reinforcement learning including RLHF, GRPO and the distinct DPO objective. All twelve chapters are expanded with numerical reasoning, independent experiments and solved exercises; accelerator performance and large pretrained-model workflows are not certified by the CPU checks.

Preprocessing and tokenization, text representation, language models, classification, sequence labeling, machine translation, generation and decoding, pretrained model families, prompting and alignment, RAG and retrieval, speech and audio, and evaluation. Organised by problem rather than by model, so the pages survive the next architecture.

Notes are two levels: a category page lists its topics, and each written topic is its own page beneath it — /notes/math/ holds /notes/math/linear-algebra/. A topic never sits beside the category it belongs to.