A free 18-module path from gradient descent to production LLMs

What it does

This is a structured, self-paced AI Engineering curriculum: 18 modules and 146+ lessons, where each lesson is a standalone blog post explaining one concept with examples, diagrams, and math where needed. It starts at supervised learning and backpropagation, climbs through Transformers, fine-tuning, RAG, and agents, and ends at inference engineering, evaluation, safety, and full system design. The author positions it as the course he wishes he’d had when starting out.

The interesting bit

The scope is unusually wide for a free resource — most curricula stop at RAG or agents, but this one goes into KV caches, paged attention, speculative decoding, and quantization, plus a module on frontier ideas like JEPA and world models. There’s also a dedicated interview-prep module, which tells you who the real audience is.

Key highlights

• 18 modules, 146+ lessons, each a detailed blog; many lessons have companion videos
• Covers the full stack: ML foundations → Transformers → fine-tuning/alignment (LoRA, RLHF, DPO, GRPO) → RAG → agents (MCP, ReAct, multi-agent) → inference (vLLM, llama.cpp, TensorRT-LLM) → evaluation and safety
• Prerequisites are just basic Python and high-school math; the needed linear algebra and calculus are explained inline
• Free to read, no sign-up, ordered so each lesson builds on the last
• Includes a glossary of key concepts and an FAQ

Caveats

• It’s a link hub, not a repo of code — the substance lives in external blog posts and videos, so depth varies lesson to lesson
• The README is upfront that it doubles as a funnel to the author’s paid live program at Outcome School
• The author notes the course “will continue to grow,” so some listed topics may still be thin

Verdict

Worth bookmarking if you’re a developer moving into AI engineering and want a single ordered reading list instead of assembling one from scattered tutorials. If you already know your MoE from your GQA, skim the module list and cherry-pick — the interview-prep and inference modules are the densest parts."