AI Engineering from Scratch is a free, open-source curriculum that runs from the math underneath modern AI up to shipping working agents, published as a repository of 523 lessons grouped into 20 phases. It is written for people who can already program but have mostly been consumers of AI tools — the sort of developer who can call a model API but has never built an attention block, a retrieval loop or an agent runtime themselves. The project frames itself around exactly that gap: the material it is promoted with cites 84 percent of students already using AI tools against 18 percent who feel ready to use them at work.
What it does
The repository is the course. It lays out an ordered path of 523 lessons across 20 phases, starting from foundations and ending at systems you can hand to other people — the tagline the author uses is "learn it, build it, ship it for others."
Two design choices set it apart from a reading list:
- Every lesson ships an artifact. The stated rule is that each lesson ends in something real — a prompt, an agent, a piece of code you can run — rather than in a quiz or a summary.
- You pick the stack. Lessons are meant to be worked in Python, Rust, TypeScript or Julia, so the concept is the assignment and the implementation language is yours. Python is the repository's dominant language and the obvious default.
The subject range, judging by the repository's own topics, covers deep learning, transformers, natural language processing, computer vision, reinforcement learning, generative AI and large language models, then moves into applied territory: AI agents, the Model Context Protocol, and swarm intelligence. It is licensed MIT, which means you can fork it, adapt it, and teach from it without asking.
The project is also multilingual. Twelve translated landing pages are committed to the repository — Spanish, French, Portuguese, German, Italian, Chinese, Japanese, Korean, Hindi, Arabic, Russian and Turkish. The README is explicit that English is canonical and that the lesson pages themselves are machine-translated on a separate translations branch, with the policy documented in the repository's i18n notes.
How it works
Everything lives in the Git repository, so the workflow is the one you already have: clone it, read the roadmap file that indexes the lessons and phases, and work through them in order. The README's contents section is the map; the roadmap file is the lesson list. A companion website, aiengineeringfromscratch.com, mirrors the material for people who would rather read in a browser than in a checkout.
Because the curriculum is language-agnostic at the implementation layer, the unit of progress is your own artifact rather than a graded submission. There is no grading, no enrollment and no instructor in the loop — the repository does not claim any of that. What it gives you is sequence and scope, which is the part most self-taught learners get wrong.
Getting started
Clone the repository, open the roadmap, and start at phase one rather than skipping to agents — the ordering is the product. Decide up front which of the four languages you will implement in and stay with it, so your artifacts accumulate into something coherent instead of four half-finished stacks. If you want a translated landing page, follow the language links at the top of the README, but treat the English lessons as the source of truth; the translated lesson text is machine-produced and lives on its own branch.
When to use it / when not
Reach for it when you want depth over speed: when you would rather implement a transformer or an agent loop than read about one, when you need a structure imposed on months of self-study, and when you want a trail of runnable artifacts you can show an employer at the end.
Skip it if you want a two-hour answer to a specific question — 523 lessons is a commitment, and this is a bad way to look something up. Skip it too if you need a credential, an instructor or feedback on your work, because none of that exists here. And note the dates: the repository was created in March 2026 and was still being pushed to in late September 2026, so it is young, moving quickly, and the material you start today may not be the material you finish.
For a self-taught engineer, the value of this repository is not any single lesson but the fact that someone has already done the sequencing: twenty phases in an order, each with something to build at the end of it. That is worth taking seriously if you have the months to spend and the discipline to work through a long path without external deadlines. If you are looking for a quick upskill, a certificate, or polished non-English course text, look elsewhere — but if you genuinely want to build the parts of an AI system yourself rather than wire together someone else's, few free resources are this ambitious in scope.