AI Engineering From Scratch Deep Dive: The Open-Source Curriculum for Building AI Systems by Hand

This repo is not a framework and not a weekend tutorial. It is a structured learning system: lesson folders, runnable code, docs, quizzes, generated website data, agent skills, prompts, scripts, and contribution rules for people who want to understand the stack below the API call.

1. ai-engineering-from-scratch in One Sentence

AI Engineering From Scratch is an open-source AI curriculum repo where each lesson aims to teach the concept, implement it from scratch, compare it with production libraries, and ship a reusable artifact such as a prompt, skill, agent, or MCP server.

2. Why It Matters

The repo matters because many AI builders can call an API but cannot explain the math, model behavior, retrieval failure, agent loop, evaluation harness, or production tradeoff underneath that call.

The curriculum's premise is straightforward: build the smaller version yourself first, then use the framework. That pattern makes PyTorch, Transformers, LangGraph, MCP, and production RAG less magical.

The strongest audience is the working engineer who wants a long path rather than a playlist. It is especially relevant for teams trying to turn AI enthusiasm into durable internal capability.

3. Architecture and Mental Model

The repository is organized around phases and lessons. Each lesson follows a consistent shape with docs, code, quizzes, and outputs, while scripts audit lessons, build catalogs, install skills, and generate the public website.

4. Smallest End-to-End Setup

The commands below are copied from the repository documentation and checked against the current research snapshot. Treat them as a starting point, then read the linked README before installing into a production environment.

git clone https://github.com/rohitg00/ai-engineering-from-scratch.git
cd ai-engineering-from-scratch

# Run a first lesson implementation
python phases/01-math-foundations/01-linear-algebra-intuition/code/vectors.py

# Install Python dependencies when needed
python -m pip install -r requirements.txt

A small first task should prove the integration before you attach it to critical data or large workspaces.

# Find your starting point inside a supported agent
/find-your-level

# Check a phase after studying
/check-understanding 3

# Inspect the generated catalog
python3 scripts/build_catalog.py --stdout

# Install lesson skills into a target skill directory
python3 scripts/install_skills.py <target-dir> --phase 14

# Validate lesson code without running heavy jobs
python3 scripts/lesson_run.py

5. Technical Deep Dive

lesson/
  docs/en.md     # explanation
  code/          # runnable implementation
  quiz.json      # check understanding
  outputs/       # prompt, skill, agent, or MCP artifact

6. Real-World Wrong vs Right Patterns

7. Common Mistakes and Current Issues

The issue tracker matters because these are young, fast-moving repos. The article uses issues as risk signals, not as proof that a project is unusable.

8. Performance, Scaling, and Cost Notes

Most early lessons are cheap to run. Later lessons involving PyTorch, Transformers, multimodal models, local inference, fine-tuning, or capstone systems can need more compute and API access.

The practical setup is to run lessons in small increments, pin a stable Python version, and avoid starting GPU-heavy or API-heavy lessons until you have read their docs and dependency expectations.

For teams, the repo's scripts and skills are useful for curriculum governance. You can assign phases, run audits, install selected skills, and keep a shared progression instead of telling everyone to browse hundreds of files.

9. Who It Is For

10. Community Signal

Web articles frame the repo as a large free AI engineering reference manual, often comparing it to a degree-style path. Some of those articles use stale lesson counts, so counts should be treated as moving metadata.

Reddit discussion is useful because it includes both excitement and skepticism: questions about AI-assisted authorship, beginner overwhelm, API costs, and whether advanced agent lessons handle reliability deeply enough.

The GitHub issue tracker shows a living curriculum: translation PRs, Rust-track requests, quiz bugs, rendering fixes, lesson wiring, and website improvements.

11. The Verdict: Is It Worth Using?

12. The Bigger Picture

This repo is part of a broader correction in AI education. After years of API-first demos, engineers increasingly need to understand data, math, model behavior, evaluation, agents, protocols, and production failure modes.

The most important habit it teaches is not any one algorithm. It is the pattern of building the small mechanism before trusting the big framework.

13. Frequently Asked Questions

14. Glossary

15. All Sources and Links

Internal Links

• Top MCP Servers
• Agentmemory Deep Dive
• Claude Code Best Practice Guide

16. Source Attribution Table