The 18-Chapter AI Compendium That Got Friends Into DeepMind and OpenAI

Henry Ndubuaku spent years filling notebooks with intuition-first explanations of maths, computing, and AI concepts. In 2025, a few friends used those notes to prep for interviews at DeepMind, OpenAI, and Nvidia. They all got in. Ndubuaku himself got into Y Combinator last year.

Now he has published the whole thing as an open-source textbook. The Maths, CS & AI Compendium runs 18 chapters, from vectors and matrices through calculus, statistics, probability, machine learning, computational linguistics, computer vision, audio and speech, multimodal learning, autonomous systems, graph neural networks, computing and OS, data structures and algorithms, production software engineering, SIMD and GPU programming, AI inference, and ML systems design.

Every chapter is marked “Available” in the repository’s outline. That is a lot of ground to cover. The compendium is not a collection of links or a reading list. It is a structured textbook written by one person, with a foreword, a study technique, and a citation entry formatted as a BibTeX reference.

The compendium also ships with an MCP server. That means any AI assistant that supports the Model Context Protocol, including Claude Code, Cursor, and VS Code, can use the compendium as a knowledge base. The server requires a local clone of the repo and comes with tools for educational purposes and example implementations. This turns the textbook into something an AI can read and reference while helping a user work through a problem.

The foreword is worth reading. Ndubuaku frames the entire project around a specific claim: that intensive, structured learning can meaningfully boost fluid intelligence. He cites Kvashchev’s experiment, a long-term Serbian study that demonstrated three years of creative problem-solving training added 10 to 15 IQ points. He compares natural IQ to quality weight initializations in a neural network, and argues that the only real advantage of a high-IQ individual is faster pattern recognition. The rest is exposure and repetition.

He then tells his own story. Between ages 3 and 10, he performed well academically without taking notes or revising. Between 11 and 13, he got cocky and dropped to the bottom half of an 80-student class. Between 14 and 15, he started reading like a normal student and finished first in his final secondary school semester. The lesson he draws is that early school curriculum works well with natural IQ, but real-world talent is powered by quality knowledge consumption and execution intensity.

The study technique he describes is straightforward. Phase one is cumulative reading: read each material after class before bed, then at the next lecture start over from the beginning until you reach the current end, filling knowledge gaps with additional research. Phase two is shadow reading before exams: read each subtitle, close the book, visualize and write an explanation for that concept, then re-read only what you missed. He compares this to masked-language modeling in machine learning.

The compendium is opinionated in its structure. It puts vectors and matrices first, then calculus, statistics, and probability, then machine learning, then domain-specific chapters for language, vision, audio, and multimodal systems, then systems and production engineering, then GPU programming and inference optimization. The ordering reflects a specific view of what a well-rounded AI researcher or engineer should know. It is not the standard university curriculum, which typically separates pure maths, applied maths, computer science, and electrical engineering into distinct departments.

The MCP server is the most interesting technical detail. Most open-source textbooks are static documents. This one is designed to be queried by an AI assistant during a coding session. That changes the relationship between the reader and the material. Instead of searching for a specific formula or concept, a user can ask their AI assistant to pull the relevant section from the compendium, get the intuition-first explanation, and apply it to the problem at hand.

The compendium is also a credential. Ndubuaku writes that his friends used these notes to prep for interviews at DeepMind, OpenAI, and Nvidia, and that they all got in and currently perform well in their roles. That is a strong signal. The textbook is not just a pedagogical exercise. It was tested against the hardest interview processes in the industry and produced results.

There are limits to what a single-author textbook can cover at this depth. Eighteen chapters spanning pure maths, applied machine learning, systems engineering, and GPU programming is a lot of territory. Some sections will inevitably be thinner than others. The compendium is also a living document, which means its quality depends on continued maintenance and updates as the field moves.

The MCP server turns the textbook into something an AI can read and reference while helping a user work through a problem.

The compendium enters a crowded field. There are excellent open-source textbooks for machine learning, including Michael Nielsen’s Neural Networks and Deep Learning, the fast.ai course materials, and the Deep Learning Book by Goodfellow, Bengio, and Courville. There are also comprehensive resources like the Stanford CS229 notes and the MIT OpenCourseWare materials. What sets Ndubuaku’s project apart is its breadth and its integration with AI assistants through the MCP protocol.

For AI builders, the compendium is a useful reference. The chapters on AI inference, GPU programming, and ML systems design cover topics that are rarely taught in university courses but are essential for shipping production models. The chapter on production software engineering includes sections on Linux, Git, codebase design, testing, CI/CD, Docker, model serving, MLOps, monitoring, and the best way to use coding agents. That is a curriculum for turning a researcher into an engineer.

The compendium is also a signal about how AI education is evolving. The best resources are no longer textbooks published by university presses. They are open-source repositories written by practitioners who have been through the interview process and built production systems. They are designed to be read alongside an AI assistant. And they are free.