AI Engineering from Scratch: 523 Free Lessons, and the 4 Phases I'd Do First
You commented COURSE, so here's the link and the order I'd do it in. AI Engineering from Scratch is a free, open-source course that lives on GitHub. It had 58,158 stars and over 10,000 forks on the GitHub API on 26 September 2026, and it was on GitHub's daily trending page the same day.
The short version: 523 lessons, 20 phases, about 342 hours, and every lesson makes you build something you keep. You don't need all of it. Do four phases, ship a project after each, and you have a portfolio a startup will actually look at.
The course: rohitg00/ai-engineering-from-scratch ~58K ★
github.com/rohitg00/ai-engineering-from-scratch
- 523 lessons across 20 phases, roughly 342 hours in total.
- Four languages: Python, TypeScript, Rust and Julia.
- Every lesson ships an artifact: a prompt, a skill, an agent or an MCP server. You finish with things you can show, not just notes.
- MIT licensed and free. No paywall, no signup.
- Hindi landing page if you prefer it: i18n/hi/README.md. English is the main version and lesson pages are machine translated.
The phases go from setup and maths, through classical ML, deep learning, transformers and LLMs from scratch, up to LLM engineering, tools and protocols, agents, multi-agent systems, production infrastructure, safety, and a final set of capstone projects.
Three ways to start
All three come from the project README. Pick one.
1. Read it on the website
Open aiengineeringfromscratch.com. No setup and no cloning. Good for your phone on the metro.
2. Let your coding agent teach you
If you already have Node.js and Claude Code or Codex installed:
npx skills add rohitg00/ai-engineering-from-scratch
Then run /start-learning in Claude Code, or pick
start-learning from /skills in Codex. A
ten-question placement quiz works out what you already know, picks your
starting phase and saves a study plan to LEARNING.md, so
every session picks up where you left off.
3. Clone it and run the code
git clone https://github.com/rohitg00/ai-engineering-from-scratch.git, then run the Phase 0 environment check
from the repo root:
python3 phases/00-setup-and-tooling/01-dev-environment/code/verify.py --route beginner
It tells you what's missing and gives you the command to fix each problem. Cloning also means the tutor can actually run each lesson's code with you instead of just reading along.
The 4 phases I'd do first
This order is for one goal: getting hired as an AI application developer. That's the role companies are hiring freshers for, and it's about building products on top of models, not training them. Skip ahead to the parts that get you there, and come back for the maths and deep learning phases once you're building.
1. Phase 0: Setup and Tooling (12 lessons)
Phase 0 covers your dev environment, Git and the tools every later lesson assumes.
Ship after: a clean GitHub profile with one repo that has a real README (what it does, a screenshot, how to run it).
2. Phase 11: LLM Engineering (17 lessons)
Phase 11 is the core of AI app work: prompt patterns, structured outputs, embeddings, RAG and advanced RAG, fine-tuning with LoRA, function calling, evaluation, caching and cost, guardrails, and building a production LLM app.
Ship after: a RAG app over something you actually care about, like your college notes or a docs site you use every day, with a live URL and a short note on how you evaluated it.
3. Phase 13: Tools and Protocols (31 lessons)
Phase 13 teaches function calling in depth, how to build MCP servers and clients, MCP security and auth, and how Agent Skills are packaged and invoked. If you only want MCP, the README has a focused 17-lesson route through it.
Ship after: an MCP server for a tool you already use, published on GitHub with install steps someone else can follow.
4. Phase 14: Agent Engineering (54 lessons)
Phase 14 builds agents from first principles: the agent loop (about 120 lines of plain Python in lesson 1), planning, reflection, tool use and memory. Lessons 31 to 46 form the agent-assisted engineering path, which is about using coding agents reliably on real repositories.
Ship after: a small agent that does one real job end to end, like triaging your GitHub issues or summarising a daily feed.
Then build one capstone
Phase 19 has 17 end-to-end projects, each 20 to 40 hours, plus deeper build tracks. Three that read well on a fresher resume:
- GitHub issue-to-PR autonomous agent
- RAG over a codebase (cross-repo semantic search)
- Stateless MCP server with a registry and governance
How I'd use it
- One phase, one shipped project. Don't start the next phase until the project has a live link.
- Post your build in public every week. That's your proof of work, and it's how founders find you.
- Take the placement quiz first so you don't redo things you already know.
- Three shipped AI projects on GitHub will get you further than finishing all 523 lessons and shipping nothing.
Want more project ideas? Read 25 projects to build to get hired in 2026, then what 2027 interviews will actually ask you.
Follow @avi_vashishta29 for one practical AI dev tool a day.