AI Engineering From Scratch: Full Course Review
/ai-engineering-from-scratch
AI Engineering From Scratch earns its name by refusing to skip the hard parts. Every algorithm gets built from raw math first, and only then does PyTorch appear. That sequence is what separates understanding from API familiarity. Worth the time if you want to know what your tools are actually doing. Skip it if you need something you can finish in a weekend.
What is AI Engineering From Scratch?
AI Engineering From Scratch is a free, MIT-licensed curriculum with 511 lessons across 20 phases taking you from linear algebra through LLM engineering, agent development, and autonomous systems. Every lesson follows a Build It / Use It / Ship It loop: implement the algorithm from raw math, see the same thing work in PyTorch, then produce a reusable artifact (prompt, skill, agent, or MCP server). Covers Python, TypeScript, Rust, and Julia. The course is designed to run on your own laptop, with EPUB and PDF editions for offline reading.
What this curriculum covers
- โ511 lessons across 20 phases from math foundations through autonomous agents and swarms
- โBuild It / Use It / Ship It loop: implement from raw math before touching frameworks like PyTorch
- โMulti-language curriculum: Python, TypeScript, Rust, Julia implementations for core algorithms
- โEvery lesson produces a reusable artifact: a prompt, skill, agent, or MCP server you can install and use
- โFocused learning paths for Model Context Protocol (17 lessons), Agent Skills (5 lessons), and Claude certification prep
- โLearning skills integrate directly with Claude Code and Codex for interactive, session-driven study
- โSix-volume book series available as EPUB and PDF for offline reading
- โPlacement quiz maps your current knowledge to a personalized starting phase and saves a plan to LEARNING.md
Why traditional AI learning falls short
Most AI material scatters the fundamentals across isolated papers, tutorials, and demos that never connect. Deploying a chatbot happens without understanding its loss curve. When you attach a function to an agent, you don't know what attention does inside the model that calls it. The problem is not a lack of resources but a lack of structure that builds upward from first principles toward production systems.
Choosing your entry point
- โขComplete beginners start with Phase 0 to set up their dev environment before touching any AI concepts
- โขPython developers who want math and machine learning foundations before jumping into deep learning frameworks
- โขEngineers building production LLM applications begin at Phase 11 to learn prompt engineering, RAG, and fine-tuning
- โขDevelopers building AI agents start at Phase 14 where the agent loop, tool use, and memory systems live
- โขTeams adopting Model Context Protocol follow the MCP path through 17 lessons covering transports, security, and registry governance
- โขThose preparing for Claude certification use the built-in certification tutor that creates a personalized study route through official exam objectives
Strengths
- โThe "build it first, then use it" structure forces you to implement attention, backprop, and tokenizers from raw math before PyTorch or sklearn ever appear โ by the time a framework shows up, you already know what it is automating.
- โEvery one of the 511 lessons ends with a portable artifact: a prompt, a skill, an agent, or an MCP server you can install and use in Claude Code, Codex, Cursor, or any compatible host immediately.
- โThe same curriculum is available as a downloadable six-volume book series (EPUB and PDF per release) and as an AI-driven tutor loop inside your terminal โ you can read it, run it, or have it teach you interactively.
- โMultiple focused entry points let you skip the full 329-hour curriculum if you already have a goal: a 17-lesson MCP path, a five-lesson Agent Skills path, or a full certification prep track for all four Claude certification levels.
What this course does not cover
- โณLimited language scope โ Python, TypeScript, Rust, and Julia only; developers in Go, Java, or other ecosystems won't find language-native implementations.
- โณNo guaranteed certification outcomes โ the README explicitly states the academy 'cannot guarantee a passing score' and is 'independent study material' not affiliated with Anthropic.
- โณBuild-from-scratch pedagogy trades speed for depth โ the raw-math-first approach leaves less time for high-level API shortcuts that matter under production deadlines.
- โณBook format is a static snapshot โ older EPUB and PDF editions remain downloadable and may lag behind the living repository curriculum.
- โณNarrow production scope โ Phase 17 covers infrastructure and deployment, but the curriculum does not extend into MLOps tooling, data engineering pipelines, or observability practices beyond its own lesson artifacts.
Other AI engineering learning resources
Frequently asked questions
Yes. It is fully open source under the MIT license, which means you can use it personally and commercially at no cost.
The full curriculum estimates around 329 hours. Focused paths let you zero in on specific areas like the Model Context Protocol or Agent Skills if you do not need the whole thing.
Some coding ability is expected. Python helps but the curriculum works with four languages, and Phase 0 covers dev environment setup from scratch.
There is no formal certificate, but the curriculum includes quizzes, a placement diagnostic, and practice exams for Claude certification preparation.
Python, TypeScript, Rust, and Julia. Each lesson provides runnable code in your language of choice.
Yes. Phase 0 assumes no prior AI knowledge and begins with development environment configuration. Your existing knowledge determines your ideal starting phase through a placement assessment.
Who should try it โ and who should skip
Pick this up if you are a developer who wants to understand AI deeply instead of just calling APIs. If you are comfortable writing code and want to know what happens inside PyTorch before you reach for it, the build-first approach is exactly what you need. It suits self-directed learners who can set their own pace. Skip it if you need quick results or want an instructor holding you accountable. Also skip it if you are already comfortable implementing backprop and attention from scratch and just need the production tooling layer.
Related repositories
Still deciding about ai-engineering-from-scratch?
One click hands the question to an AI along with this page โ see what it says about ai-engineering-from-scratch.