AI Engineering from Scratch: 523 Hands-On Lessons with AI Tutor Integration

The Problem: Using AI vs. Understanding AI

84% of students already use AI tools like ChatGPT and Claude, yet only 18% feel ready to apply them professionally. This gap exists because most resources either "feed knowledge" (courses, videos, papers) that leave learners unable to build, or "give tools" (frameworks, libraries) that hide the underlying mechanics. The project ai-engineering-from-scratch (rohitg00/ai-engineering-from-scratch) aims to break this dichotomy with the motto "Learn it. Build it. Ship it for others."

Project Overview

Created in March 2026, the repository has gained 52k+ stars in 8 months and attracts 110k readers per month. It is MIT-licensed, written primarily in Python with TypeScript, Rust, and Julia, and continuously updated. The author, Rohit, is also the creator of the Agent Memory project, lending practical infrastructure experience to the curriculum design.

Curriculum Structure: 523 Lessons, 20 Phases, ~342 Hours

The curriculum spans five layers organized as a directed acyclic graph (DAG) allowing branched specialization:

Foundations (Phases 0-2) : Development environment, math intuition via code, classical ML.

Deep Learning Core (Phases 3-6) : Neural networks, computer vision, NLP, audio. Principle: "No frameworks until you build one." Phase 3 constructs a neural network from first principles with zero dependencies.

Language & Generation (Phases 7-10) : Transformer architecture, generative AI, reinforcement learning, and building and training an LLM from scratch (Phase 10, 24 lessons) .

Application Engineering (Phases 11-13) : LLM engineering, multimodal, tools & protocols (MCP, Agent Skills).

Agents & Production (Phases 14-19) : Hand-written agent loops, autonomous systems, multi-agent clusters, production infrastructure, model alignment, capstone project.

Learners can follow the full path or branch into focused tracks (e.g., vision vs. language after Phase 6).

Unique Mechanism: Every Lesson Delivers a Reusable Tool

Unlike courses that end with "congratulations, you learned X," each lesson here produces a concrete artifact — a prompt, skill, agent, or MCP server — that integrates into daily work. For example, Phase 14 Lesson 1 teaches the agent loop in ~120 lines of pure Python (zero dependencies). The deliverables are agent_loop.py, an agent-loop skill, and a prompt-debug-agent prompt. Completing the full course yields 523 personally built, understood, and reusable artifacts, creating a compounding loop: building ensures understanding; reusing reinforces it.

AI Tutor Integration: Turn Your Coding Agent into a Personal Instructor

A single command installs the curriculum into Codex, Claude Code, Cursor, etc.:

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

The agent then drives a three-step loop:

Placement : A 10-question quiz maps existing knowledge to a starting phase, generating a personalized LEARNING.md plan.

Learning Loop : Each lesson follows concept → math derivation → hands-on coding → quiz, with the agent guiding and tracking progress.

On-Demand Help : The course-guide skill routes questions (e.g., "explain attention" or "my loss is NaN") directly to the relevant lesson.

This transforms passive video watching into interactive, paced mentorship.

Focused Paths for Specific Goals

MCP Only : 17-lesson track covering stateless requests, transport, bidirectional workflows, security, reliability, registry governance, consistency verification; produces MCP-LEARNING.md with experimental evidence.

Agent Skills Only : 5-lesson track on contracts, discovery, invocation, sandbox boundaries, publishing/evaluation for portable Agent Skills.

Claude Certification Prep : Open-source Claude Certification Academy aligning with Anthropic's four official tracks (Associate Foundations, Developer, Architect, Architect Professional), each with mapped lessons, runnable labs, diagnostic questions, and full mock exams.

The full 20-phase curriculum also compiles into 6 EPUB/PDF volumes released with each update for offline study.

Value Proposition: Three Combined Differentiators

From-Scratch Depth : Not API tutorials; learners internalize how neural networks, Transformers, and LLMs actually work.

Compounding Artifacts : Every lesson yields a production-grade tool the learner authored, turning "knowledge in brain" into "tools in toolbox."

Personalized AI Tutoring : Coding agents adapt to the learner's level, pace, and questions, replacing passive consumption with guided construction.

Individually these exist elsewhere; combined they form a closed loop rare in open-source education.

Target Audience & Caveats

Suitable for : Developers wanting deep AI principles beyond API calls; those willing to write code, not just watch videos; coding-agent users wanting to repurpose them as tutors; engineers entering Agent/MCP/LLM engineering systematically.

Not suitable for : Seekers of "5-minute AI tool setup"; non-coders; those expecting immediate high-paying jobs — the course builds a solid foundation, but real-world application remains the learner's responsibility.

The project is completely free (MIT) with simplified Chinese and other language support. Even as a reference "AI engineering dictionary" it warrants a bookmark.

Signal: A Shift in AI Learning Paradigm

The project's rapid adoption (52k stars, 110k monthly readers in 8 months) signals a broader transition: AI-era education is moving from "passive absorption" to "hands-on construction + AI as personal tutor." Courses are becoming dynamic learning operating systems that plug into coding agents, adapt to individual levels, and emit tangible outcomes at every step — a likely direction for AI engineering education.