I Found a Free GitHub Repo With 435 AI Engineering Lessons and It Feels Like a Full Degree

AI ENGINEERING

Most people learning AI today start from the wrong end.

They open an LLM API, copy a few lines of Python, build a chatbot, connect a vector database, and suddenly feel like they are doing AI engineering. And to be fair, they are building something useful. But there is one uncomfortable problem: many people can make AI tools work without actually understanding why they work.

That gap becomes obvious very quickly.

You can call an embedding model, but do you understand what vectors are doing? You can fine-tune a model, but can you explain loss curves, backpropagation, tokenization, attention, or evaluation? You can build an agent with tools, but do you really understand the agent loop, planning, memory, failures, or production risks?

That is why I found rohitg00/ai-engineering-from-scratch so interesting.

GitHub repo: https://github.com/rohitg00/ai-engineering-from-scratch

This is not another “build a chatbot in 10 minutes” tutorial. It is a full open-source AI engineering curriculum with 435 lessons across 20 phases, covering everything from math foundations to machine learning, deep learning, transformers, LLMs, agents, multimodal AI, production systems, safety, and capstone projects. The project describes itself as a curriculum where every algorithm is built from raw math first, before using production frameworks.

And that is what makes it different.

The Problem With Most AI Learning Today

AI education has become very scattered.

One day you read a blog post about RAG. The next day you watch a video about transformers. Then someone shares a prompt engineering guide. Then another person says agents are the future. Then someone else says you should learn MCP, multimodal models, fine-tuning, evals, RLHF, safety, and deployment.

All of these topics matter, but they often come in pieces.

The result is that many learners collect fragments without building a real mental model. They know tools, but not foundations. They know workflows, but not first principles. They know how to use models, but not how models think, fail, learn, or interact with systems.

AI Engineering from Scratch tries to solve that by giving the learning path a spine.

It starts from the bottom: setup, tooling, linear algebra, calculus, probability, optimization, and classical machine learning. Then it moves into deep learning, computer vision, NLP, transformers, generative AI, reinforcement learning, LLMs, multimodal systems, tools, protocols, agents, autonomous systems, swarms, infrastructure, safety, and final projects.

In simple words: it does not just teach you how to use AI. It teaches you how AI is built.

What This Repo Actually Is

At its core, this repository is a structured AI engineering manual.

It has:

20 phases 435 lessons Around 320 hours of learning Four programming languages: Python, TypeScript, Rust, and Julia Reusable outputs from lessons: prompts, skills, agents, and MCP servers

The official website also presents the same idea clearly: 435 lessons, 20 phases, and every algorithm built from raw math before importing a framework.

That “from scratch” part is important.

For example, instead of directly jumping into PyTorch or Transformers, the curriculum pushes you to understand the smaller pieces first. You learn what vectors are. You implement gradient descent. You understand loss functions. You build tokenizers. You study attention. You build the agent loop. Then when you use a framework, it does not feel like magic anymore.

This is a much deeper learning style.

It is slower than copying code from a tutorial, but it builds real confidence. And in AI engineering, confidence matters because production AI systems break in strange ways. Models hallucinate. Retrieval fails. Agents loop forever. Evaluation metrics lie. Prompt changes create unexpected behavior. A system that works in a demo may collapse with real users.

To debug these things, you need more than API knowledge. You need understanding.

Why This Repo Is Useful

The biggest value of this repo is that it connects theory with building.

Many AI courses are either too theoretical or too shallow. Some explain math but never help you ship anything. Others help you build quick projects but hide the core ideas behind libraries. This repo tries to sit in the middle.

Each lesson follows a practical loop: understand the problem, derive the idea, write the code, run tests, and keep the artifact. The project also says every lesson ships something reusable, such as a prompt, a skill, an agent, or an MCP server.

That means the repo is not just for reading. It is for building a portfolio of working AI components.

By the end, you are not only saying, “I completed a course.” You can say, “I built pieces of AI systems myself.”

That is powerful for students, developers, data scientists, ML engineers, and anyone trying to move from AI user to AI builder.

The 20-Phase Structure

The curriculum is designed like a ladder.

It begins with Phase 0: Setup & Tooling, where you prepare your environment. Then it moves into Math Foundations, ML Fundamentals, and Deep Learning Core. After that, it expands into specialized areas such as computer vision, NLP, speech, transformers, generative AI, reinforcement learning, and LLMs.

The later phases are especially relevant for today’s AI market: LLM engineering, multimodal AI, tools and protocols, agent engineering, autonomous systems, multi-agent swarms, infrastructure, production, ethics, safety, alignment, and capstone projects.

The roadmap shows estimated time for each phase and tracks progress across the curriculum. It lists a total estimated time of around 314 hours, while the README describes the curriculum as roughly 320 hours.

That is not a weekend course.

And that is actually the point.

This is for people who want to take AI seriously.

The Best Part: It Teaches Agents From First Principles

One reason this repo feels especially relevant now is its focus on agents.

Today, many developers are trying to build AI agents using frameworks. They use LangChain, CrewAI, AutoGen, OpenAI tools, Claude tools, MCP servers, and other systems. But when something fails, they often do not know whether the problem is the prompt, tool schema, memory, planning logic, retrieval, model limitation, or orchestration.

This repo includes a large Agent Engineering section. According to the README details, Phase 14 alone contains 42 lessons. It covers agent loops, memory, planning, frameworks, benchmarks, production ideas, and workbench-style tooling.

That is extremely useful because agents are not just “chatbots with tools.” They are systems that reason, act, observe, recover, and continue. If you understand the loop, you can debug the system better.

How to Use It

The easiest way is to open the website and read the lessons directly. The second way is to clone the repo and run the code locally.

Basic setup:

git clone https://github.com/rohitg00/ai-engineering-from-scratch.git
cd ai-engineering-from-scratch
python phases/01-math-foundations/01-linear-algebra-intuition/code/vectors.py

You can also install the course skills into an AI coding assistant using the included scripts. The README explains that the repository ships many skills and prompts under the lesson outputs, and the install script can copy them into a target directory for tools like Claude, Cursor, Codex, OpenClaw, Hermes, or similar agent environments.

That is a clever idea.

The curriculum does not just teach AI concepts. It also creates tools that can help you learn and work faster.

Why It Is Getting Attention

AI is moving fast, but the people who understand foundations will have a long-term advantage.

Anyone can call an API. Fewer people can explain what happens under the hood. Even fewer can build reliable AI systems, debug them, evaluate them, and ship them safely.

That is why a repo like this stands out.

It is open source, MIT licensed, practical, deep, and structured. It does not promise shortcuts. It gives a long path, but the path makes sense.

The repository has also gained strong public attention, with GitHub search results showing thousands of stars and forks. The GitHub issue page currently shows around 16.5k stars and 2.9k forks, though those numbers may keep changing as the project grows.

Final Thoughts

What I like most about AI Engineering from Scratch is its mindset.

It does not treat AI as a magic API. It treats AI as an engineering discipline.

That matters because the next generation of AI builders will not be judged only by how many tools they know. They will be judged by whether they can understand systems deeply, build reliable products, debug failures, and make intelligent design choices.

Sources

GitHub repo: https://github.com/rohitg00/ai-engineering-from-scratch Official website: https://aiengineeringfromscratch.com/

A quick note before you go 👋

I break down real AI shifts before they hit the mainstream 🚀 Click Follow now so you do not miss what matters next and drop a clap 👏 if this helped.