AI Engineering from Scratch

What it is

AI Engineering from Scratch is a free course by Rohit Ghumare, kept as a GitHub repository under the MIT license. Its README opens with the job: implement model internals, retrieval pipelines and agent runtimes, test them, inspect failures, and keep the code and evaluation results. The lessons sit in twenty phases, from setup, math and classical machine learning through deep learning, transformers and LLMs, to tools and protocols, agents, production and capstone projects. Code runs in Python, TypeScript, Rust and Julia. Every lesson follows the same six beats: motto, problem, concept, Build It, Use It and Ship It. You write the small version first, then run the same thing through a production library. The lesson ends with something you keep, such as a prompt, a skill, an agent or an MCP server.

1. Pick a start point: the website, a tutor or a clone
2. Check your tools with the Dev Environment preflight
3. Read a lesson’s docs and run its code
4. Keep the command, the output and what you changed
5. Ship the lesson’s prompt, skill, agent or MCP server

Who it suits

Anyone who can write code in any language and wants to understand how AI works, not just how to call an API. The README’s prerequisites say exactly that, and add that Python helps. You do not have to start at the beginning. Its table sends a Python developer to Phase 1, someone who knows ML to Phase 3, and a senior engineer who only wants agents to Phase 14.

Skip it if all you need is to call a model API from an app. The README says the course is for people who want to understand how AI works, not just call APIs.

Where to use it

The lessons are Markdown, code and quizzes, so you can read them on the website, learn them through a coding agent, or run them on your own machine.

• The website: aiengineeringfromscratch.com, with nothing to install
• A coding agent: the README names Claude Code, Codex and Cursor, and any agent that reads SKILL.md, as hosts for the tutor skills
• Your own computer: a clone with Python 3, plus Node.js for the TypeScript lessons and Rust or Julia for the lessons in those languages
• Cloud GPUs: the Dev Environment lesson names Google Colab and cloud GPUs for the training-heavy lessons

Most lessons work on a CPU, the Dev Environment lesson says.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with the steps from the maker’s pages. The website needs nothing at all. For the tutor you install skills with npx, and for lesson code you clone the repository.

1. Run node --version, npx --version and python3 --version to see what you have.
2. Run the install line below, then pick your agent and where the skills go when the installer asks.
3. Check that your agent finds the skills in that folder, as the README says to.
4. Clone the repository when you want lesson code to run, by you or by the tutor. Without a clone the learn skill traces code by hand, and the certification guides ask for a clone so the tutor can run every lab and test.

npx skills add rohitg00/ai-engineering-from-scratch

git clone https://github.com/rohitg00/ai-engineering-from-scratch.git
cd ai-engineering-from-scratch
python3 phases/00-setup-and-tooling/01-dev-environment/code/verify.py --route beginner

How to use it

Start the tutor by name. The README gives each host its own form, and the skills say never to show a Claude Code slash command as universal syntax.

In another host, ask in words, as in “Use start-learning to begin the course.”

start-learning asks why you are learning, how much time you have and what you want to build. It then runs a ten-question placement quiz and writes LEARNING.md in your current folder. After that, learn teaches one lesson per session: concept, math, code, quiz. The README’s way of working through any lesson:

1. Read docs/en.md and say the core idea in your own words.
2. Type and build the important code.
3. Run the lesson command from the repository root.
4. Keep the evidence: the command, the folder, the exit code, the output and the artifact.
5. Move on once you can explain the output and change one small thing without guessing.

Example

You are a Python developer with a free evening. You clone the repository, run the preflight, then the first math lesson:

python3 phases/00-setup-and-tooling/01-dev-environment/code/verify.py --route beginner
python3 phases/01-math-foundations/01-linear-algebra-intuition/code/vectors.py

The preflight separates what you need now from what can wait, and prints a corrective command for anything required that is missing. The vectors script has no dependencies. It ends by showing that a matrix times a vector is the operation inside a neural network layer. Save that output, since the README calls it your first evidence. Then open the lesson’s docs/en.md and read what the script just showed you.

For newcomers and experts

If you are new. A phase is a group of lessons on one topic, and phases 0 to 19 stack from setup up to capstone projects. A skill is a Markdown file called SKILL.md that tells a coding agent how to do one job, here how to tutor you. MCP, the Model Context Protocol, is a standard for connecting AI tools to other software. Phase 0 sets up your computer, and the preflight tells you what is missing and how to fix it. If installing a tutor feels like too much, open the first lesson on the website and just read.

For experts. Each tutor keeps its state in a file in your working folder: LEARNING.md, MCP-LEARNING.md, AGENT-SKILLS-LEARNING.md, REVERSE-ENGINEERING-LEARNING.md, CLAUDE-CERTIFICATION.md and MCPA-CERTIFICATION.md. The focused routes follow JSON manifests under learning-paths, not the numeric order of Phase 13. The MCP route targets the 2026-07-28 stateless protocol revision, where each request carries its own protocol version and client capabilities in _meta, with no handshake first.

The preflight has a route for each goal: beginner, ml-foundations, llm-engineering, agents, mcp, agent-skills and certification. The three projects ship graders you run with scripts/project_test.py and --strict. scripts/install_skills.py installs the artifacts the lessons produce in a skills, by-phase or flat layout, and scripts/scaffold_workbench.py drops the Phase 14 agent workbench pack into another repository.

Good to know

The website needs no install. The tutor runs inside your own coding agent, and the lesson code runs on your own machine with python3 and the other runtimes.

• Cost: the course is free under MIT. The pages give no price for the coding agent or model that runs the tutor, so check your own plan
• Your data: tutor progress stays in Markdown files in your working folder. Without a clone, the learn skill fetches lessons from raw.githubusercontent.com, and the pages say nothing about what your coding agent’s provider keeps, so check its own terms
• Keys: the certification labs run locally with no key by default, and the guide says never to put an API key in source, a prompt or a state file. Phase 0 has an APIs & Keys lesson that these notes did not read
• Counts move: the README gave 542 lessons on 3 November 2026 and the Dev Environment lesson says 500+, so check the README for the current count
• Pages disagree: the setup preflight asks for Python 3.11 or later, while two of the three projects, scripts/lesson_run.py and the Reverse Engineering tutor say 3.10 or later
• Reverse Engineering: the tutor covers programs you built or have written authorization to analyze, and declines protection bypass, DRM, anti-cheat and software you do not own
• Certifications: independent study material, not affiliated with Anthropic, the Agentic AI Foundation or the Linux Foundation. Practice scores are not official and promise no pass
• Removal: the pages give no uninstall steps. Delete the clone and the skill folders the installer wrote
• License: MIT. The LICENSE file reads “Copyright (c) 2026 Rohit Ghumare”