AI Engineering From Scratch (52K★): Build It → Use It → Ship It
Post by Sandhya Ahuja on LinkedIn (2026-09-05) — free curriculum that teaches AI engineering by building the mechanism yourself before using the production abstraction.
The Post (full text)
A 52K-star GitHub repo is trying to teach AI engineering in a way most courses never do => build the mechanism yourself first, then use the production abstraction, then ship something with it.
AI Engineering From Scratch is a free curriculum with 523 lessons across 20 phases, moving from tooling, linear algebra and classical ML into deep learning, transformers, LLMs, multimodal systems, MCP, agent engineering, autonomous systems, swarms, infrastructure, safety and capstone projects.
The lessons follow a consistent Build It → Use It → Ship It loop. You might implement the underlying idea with raw math or minimal code, then reproduce it with the framework people actually use, and finish by producing a reusable artifact such as a prompt, SKILL .md, agent, or MCP server.
The repo also includes runnable implementations across Python, TypeScript, Rust and Julia, placement and tutoring skills that can track progress in files like LEARNING.md, focused learning routes for MCP and Agent Skills, and a six-volume book build generated from the same lesson sources.
The curriculum is organized around understanding the layer beneath the abstraction before depending on the abstraction — which is exactly the gap many AI engineers hit when they can assemble systems but struggle to explain why those systems work.
Comments
- Michael Aksoy: “This isn’t AI Engineering, its still missing core and fundamental principles of AI.”
Key Takeaways
- Free, MIT-licensed, 52K+ stars — 523 lessons / 20 phases (~320h), linear algebra → swarms → safety → capstones.
- Pedagogy: write the backprop, tokenizer, attention mechanism, and agent loop by hand before any framework gets imported.
- Build It → Use It → Ship It loop: raw math/minimal code → real framework → reusable artifact (prompt, SKILL .md, agent, MCP server).
- 4 languages: Python, TypeScript, Rust, Julia.
- Extras: LEARNING.md progress tracking, MCP + Agent Skills learning routes, 6-volume book build from the same lesson sources.