If you wanted to learn AI engineering from the ground up without paying for a bootcamp, this is the repo. 27,000 stars, 485 lessons across 20 phases, completely free under MIT.
Repo: github.com/rohitg00/ai-engineering-from-scratch
What it covers
The curriculum runs from math foundations (linear algebra, calculus, probability) all the way to autonomous multi-agent systems. 20 phases, roughly 320 hours of content. Every lesson follows the same six steps: read the problem, derive the math, write the code from scratch, then rebuild it in a real framework.
You implement backpropagation, tokenizers, and attention mechanisms by hand before you ever touch PyTorch. Four languages throughout: Python, TypeScript, Rust, and Julia.
Key phases
- Phases 1-3: Math foundations, ML fundamentals, deep learning core
- Phases 4-6: Computer vision (28 lessons), NLP (29 lessons), speech and audio
- Phase 10: Build an entire LLM from scratch (tokenizer, pre-training, RLHF, DPO, quantization)
- Phase 14: Agent engineering (42 lessons, the largest phase). LangGraph, CrewAI, Claude Agent SDK, MCP
- Phase 16: Multi-agent systems and swarms
- Phase 18: Ethics, safety, alignment, red-teaming
- Phase 19: 17 end-to-end capstone projects
Find your level
The companion site has a placement quiz that maps you to the right starting phase based on what you already know. Recommended pace is 1-2 lessons a day, fundamentals first.
Site: aiengineeringfromscratch.com
What you get out of it
Every completed lesson produces a reusable artifact: prompts, installable agent skills, MCP servers, or working code you can put on your CV. By the end you have 485+ portfolio pieces you actually understand, not just certificates.
The tradeoff
320 hours with no instructor, no deadlines, no cohort. "From scratch" means from scratch. The first three phases are pure math before you write a single line of ML code. If you can't self-motivate through linear algebra without a deadline, you'll stall early.