Build AI systems from model internals to agent deployment with this free code-based curriculum
Learn to implement model internals, assemble retrieval pipelines, and test agent runtimes in one place. You can start learning by reading on the website, conversing with a coding agent, or running code locally.
A curriculum built on 20 phases
The course consists of 20 phases, starting from mathematics and ending with agents and production. Each phase builds on the previous one, allowing you to skip lower levels if you already know the material. However, skipping foundations makes it harder to diagnose issues in later stages, so choosing the right starting point based on your background is important. Lessons are implemented in various languages, including Python, TypeScript, Rust, and Julia, and all content is available for free under the MIT license.
Learning through practical projects
Beyond reading concepts, the course provides three projects equipped with step-by-step starter code, reference implementations, and local graders. In the retrieval evaluation lab, you build a comparison tool to check if the NDCG metric improves and detect regressions. The agent trace debugger generates reports that separate and visualize duplicate token usage through JSONL parsing and interval arithmetic. The tool call firewall project uses Rust and Python to verify the caller's role and path and implement a single-use approval mechanism. Each project is initialized with the python3 scripts/project_test.py command and progresses by passing tests at each step.
Using a coding agent as a tutor
If you have an environment with Node.js and npx installed, you can use a coding agent as a personal tutor. After installing with the npx skills add rohitg00/ai-engineering-from-scratch command, enter start-learning or /start-learning depending on the host, such as Codex or Claude Code. A 10-question batch quiz diagnoses your current level, and a personalized learning plan is saved to LEARNING.md. Subsequently, the learn skill facilitates interactive learning one lesson per session, following the order of concepts, mathematics, code, and quizzes. If you want to focus on specific topics, dedicated paths of 17 or 5 lessons for MCP or agent skills are also provided.
Environment and time requirements before starting
This course is suitable for those with coding experience who want to understand how AI works. If you are new to programming, start from Phase 0 for approximately 306 hours; if you know Python, start from Phase 1 for about 270 hours; if you are a senior engineer, start from Phase 14 on agent engineering for about 60 hours. Running code locally requires Python 3.10 or higher and Git, while using the agent tutor requires Node.js 22.18 or higher and npx. All lessons are designed to run commands from the repository root directory, so it is advisable to check your working directory and dependency status in advance.
