AI Engineering from Scratch is a free, open-source (MIT licensed) AI engineering learning course repository, located at https://github.com/rohitg00/ai-engineering-from-scratch，配套网站为 aiengineeringfromscratch.com. The course slogan is "Learn it. Build it. Ship it for others."
[Purpose and Positioning] The course is aimed at learners who want to understand how AI actually works, rather than just calling APIs, and is also suitable for people who already know how to write code (in any language, Python preferred). The repository covers 528 lessons across 20 stages, supports four languages: Python, TypeScript, Rust, and Julia, and provides versions in more than 30 languages, including English, Simplified Chinese, Traditional Chinese, Japanese, and Korean. The course provides career path comparisons, prerequisites, and explanations of study duration.
[Main Capabilities and Content] The course is organized in a stacked manner by stage: mathematics is the foundation, and agents and production are the roof. The core exercise is the "Build It / Use It" structure—first implement the algorithm yourself by hand, then run the same functionality through a production library. Each stage contains a lesson list; for example, Stage 0 is "Setup & Tooling," with 13 lessons in total. The repository also includes several deep build series (multi-lesson series courses that build complete subsystems from scratch).
Key projects and hands-on practice: The repository provides three projects with staged starter code, reference implementations, and local graders. The first is a retrieval ranking validation project, using Python 3.10+, first reviewing RAG and model evaluation, then implementing ranking validation, precision and recall, rank-sensitive metrics, and system comparison, preserving reproducible comparisons, per-query differences, and the judgments used for scoring. The second is a distributed tracing performance project, using Node.js 22.18+ and Python 3, implementing JSONL parsing, parent validation, interval computation, and an inspectable timeline, preserving input traces, an HTML timeline, and a JSON regression report. The third is a tool-call approval project, using Rust and Python 3.10+, reviewing tool schema design and security boundaries, validating call envelopes, checking caller roles and paths, and consuming approvals bound to the exact request and content, preserving audit receipts that show the requested action and policy decision.
How to learn: You can open completed lessons directly on the website, with no installation or cloning required; you can also learn with an AI tutor. If Node.js, npx, and a coding agent that supports skills are already installed, you can use the installer to turn the coding agent into a tutor; after choosing the host and scope, use start-learning in Codex, and in Clauduse /start-learning etc. in Code. A ten-question placement quiz can map existing knowledge to a starting stage and save a personalized learning plan to LEARNING.md. There are also the learn and course-guide skills: learn explains one lesson at a time (concepts, math, code, quiz), while course-guide can jump directly to the course corresponding to the knowledge point where you are stuck. When you only need Model Context Protocol (MCP), you can use the corresponding host to invoke it, create MCP-LEARNING.md, and follow a 17-section path to learn about stateless requests, transport, bidirectional work, security, reliability, registry governance, and consistency evidence. When you only need Agent Skills, you can follow a five-section path to learn about contracts, discovery, invocation, sandbox boundaries, release evaluation, and real-host portability.
Artifacts and tools: each course produces reusable artifacts. The repository provides 400 skills and 100 prompts under phases/**/outputs/, which can be installed via scripts/install_skills.py (requires cloning the repository), and supports tag filtering, dry runs, and installation by agent layout; by default it refuses to overwrite existing targets, use --dry-run to preview conflicts or --force to overwrite, and writes manifest.json. The capstone in phase 14 provides a reusable Agent Workbench toolkit (AGENTS.md, schemas, init / verify / handoff scripts), which can be scaffolded into any repository with a script, yielding seven workbench surfaces, a starting task_board.json, and a fresh agent_state.json.
Books and certification: the 20-stage core curriculum under phases/ is compiled into a six-volume book, with EPUB and PDF built by CI from the same course source and attached to the GitHub release; the book is a snapshot, while the repository is the live version. The Claude Certification Academy is a free and open-source exam preparation project for the four official Claude certification tracks, providing blueprint-mapped courses, runnable labs, diagnostics, a capstone, and original mock exams; the MCPA certification course is aimed at the Agentic AI Foundation's Model Context Protocol Associate exam, and includes 34 lessons, covering five exam domains, and provides diagnostics, a capstone, and three sets of original mock exams. Both certifications are independent study materials based on public exam objectives, are not affiliated with the relevant organizations, and do not guarantee a pass.
[Applicable scenarios and limitations] Applicable: developers who want to systematically understand AI fundamentals, retrieval pipelines, and agent runtimes; practitioners who want to produce reusable tools while learning; people preparing for relevant certification exams. Limitations: the repository does not provide information beyond fees, prices, and specific platform requirements; both the certification academy and the MCPA course state that they are not affiliated with the relevant official organizations, do not reproduce real exam questions, and do not guarantee passing scores; some learning paths (such as real-host discovery, invocation, scripts, and uninstall evidence) remain pending until preflight is available. Other details not explained in the materials are not currently provided.
