Build AI from raw math before you’re allowed to use PyTorch

What it does

This is a free, open-source curriculum that teaches AI engineering end-to-end across 20 phases, from linear algebra to autonomous agent swarms. Every lesson follows the same loop: derive the math, build the algorithm by hand in raw Python (or TypeScript, Rust, or Julia), then run it again through a production framework like PyTorch. Each phase ships reusable artifacts—prompts, agent skills, and MCP servers—meant to be pasted into real workflows, not just filed away as homework.

The interesting bit

The pedagogy is deliberately inverted: frameworks appear only after you have written the smaller version yourself. The curriculum also dogfoods its own material by shipping agent skills—such as a placement quiz and per-phase understanding checks—that plug directly into Claude, Cursor, or Codex.

Key highlights

• 20 stacked phases covering math, deep learning, transformers, LLM engineering, agents, and production infrastructure
• Every lesson produces a tangible artifact: prompts, skills, autonomous agents, or MCP servers
• ~320 hours of material, MIT licensed, designed to run on a laptop
• Four languages supported: Python, TypeScript, Rust, and Julia
• No video content or hand-holding; each lesson is a self-contained folder with narrative docs and runnable code

Caveats

• The provided sources show conflicting lesson counts (503 versus 473), suggesting the content is still expanding or shifting.
• Only the first phase is fully expanded in the README; the depth of completed material for later phases is not visible in the truncated source.

Verdict

Worth bookmarking if you are a developer with patience and a genuine desire to understand AI internals rather than just wire APIs. Skip it if you need a five-minute deploy script or a weekend crash course.