What it does This repo is a complete, self-paced AI engineering curriculum spanning 20 phases and roughly 320 hours. It starts with linear algebra and ends with autonomous agent swarms, covering deep learning, transformers, LLMs, multimodal models, RL, and production infrastructure along the way. Every lesson follows the same rigid loop: derive the math, code it raw, then redo it in PyTorch or sklearn. No videos, no copy-paste deploys.

The interesting bit The “Ship It” beat is the twist. Each lesson doesn’t just teach — it produces a reusable artifact you can actually use: prompts, agent skills, or MCP servers. You don’t finish with a certificate; you finish with a portfolio of 503 tools you built and therefore understand. There’s even a built-in placement quiz (/find-your-level) that maps your knowledge to a starting phase with hour estimates.

Key highlights

• Four languages: Python, TypeScript, Rust, Julia — not just Python comfort food
• Every algorithm built from scratch first: backprop, tokenizer, attention, agent loop
• Structured lesson format: code/, docs/en.md, outputs/ (prompts/skills/agents/MCP servers)
• Built-in agent skills for Claude, Cursor, Codex, etc. with SKILL.md files
• Free, MIT-licensed, designed to run on your own laptop

Caveats

• The README claims 503 lessons in one place and 473 in another; the exact completed count is unclear
• “~320 hours” is an estimate, not a measured figure
• Most lessons appear to be text/docs; the actual code completion status across all 20 phases isn’t visible in the truncated README

Verdict For the developer who suspects they’re just “calling APIs and hoping” — this is the antidote. Skip it if you want quick tutorials or video explanations; it demands you write code and derive equations.