Build your own online course platform! Self-hosted, pay once and own it forever — with AI you can draft course content and outlines in minutes, and keep 100% of your revenue.
AI Engineering from Scratch: The 49K-Star Open Source Curriculum That Teaches You to Build AI, Not Just Use It
In March 2026, a London-based DevRel named Rohit Ghumare published an open-source curriculum called ai-engineering-from-scratch. Five months later, it had crossed 49,000 GitHub stars — one of the fastest-growing AI learning resources ever created.
The opening stat explains why it resonated: 84% of students already use AI tools. Only 18% feel prepared to use them professionally.
This curriculum is the attempt to close that gap — with 511 lessons, zero paywalls, and a method that makes you build the algorithm before you’re allowed to use the library.
The curriculum: a 20-phase stack from math to swarms
The 511 lessons (≈329 hours) are organized as a progressive stack — math at the floor, autonomous agent swarms at the roof:
| Phase | Focus | Lessons |
|---|---|---|
| 0 | Setup & Tooling (Git, GPU cloud, Docker, Jupyter) | 12 |
| 1 | Math Foundations (linear algebra, SVD, Fourier, Bayes) | 22 |
| 2 | ML Fundamentals (regression, SVMs, KNN, time series) | 18 |
| 3 | Deep Learning Core (backprop from scratch) | 13 |
| 4 | Computer Vision (CNNs, YOLO, Stable Diffusion, NeRF, SAM 3) | 28 |
| 5 | NLP (spaCy, NLTK, Hugging Face) | 29 |
| 6 | Speech & Audio (Librosa, Whisper) | 17 |
| 7 | Transformers (attention from masked averages) | 14 |
| 8 | Generative AI (VAEs, GANs, Latent Diffusion, ControlNet) | 15 |
| 9 | Reinforcement Learning (Q-Learning, PPO, RLHF reward modeling) | 12 |
| 10 | LLMs from Scratch (pretrain a 124M GPT!) | 24 |
| 11 | LLM Engineering (prompting, RAG, LoRA, LangGraph) | 17 |
| 12 | Multimodal AI | 25 |
| 13 | Tools & Protocols (MCP, agent skills) | 23 |
| 14 | Agent Engineering (ReWOO, Reflexion, MemGPT, voice agents) | 42 |
| 15 | Autonomous Systems (long-horizon agents, STaR, alignment) | 22 |
| 16 | Multi-Agent & Swarms (A2A, Society of Mind, BFT) | 25 |
| 17 | Infrastructure & Production (vLLM, FP8, gateways, SOC 2) | 28 |
| 18 | Ethics, Safety & Alignment (Goodhart, jailbreaks, watermarking) | 30 |
| 19 | Capstone (17 full products, 9 deep-build tracks) | 85 |
The endgame is Phase 19: 17 full-scale products (terminal-native coding agent, codebase RAG, real-time voice assistant) and 9 deep-build tracks (BPE tokenizer, distributed training).
The method: Build It, Then Use It
The signature pedagogy kills black-box magic:
- Build It — implement the raw math using only Python/NumPy and the standard library. No frameworks allowed. You write backpropagation by hand before you ever touch PyTorch.
- Use It — implement the same algorithm with industry libraries (PyTorch, scikit-learn, Hugging Face) and diff the outputs to see what the framework does under the hood.
Every lesson follows the same six-beat loop: Motto → Problem → Concept → Build It → Use It → Ship It.
And Ship It is literal: each lesson produces a reusable artifact. The full repo yields 511 assets — 396 skills, 99 prompts, agents, and MCP servers — that you can load directly into your AI coding assistant. You graduate with a portfolio, not a certificate.
Evidence-based learning is enforced: you keep execution proof (commands, exit codes, meaningful output) as you go, and only continue when you can explain the output and make one small change without guessing.
The creator: Rohit Ghumare
Ghumare isn’t a random course seller — he’s a heavyweight in the agent-infrastructure world:
- Head of DevRel at Motia.dev, based in London
- CNCF Ambassador, CNCF 2025 Marketing Chair, Platform Engineering Ambassador
- Docker Captain, Google Developer Expert, AWS Community Builder (5 years)
- Claude Ambassador and Devin Ambassador
- Creator of Agent Memory (iii-engine): 27k stars, 72k monthly downloads — persistent memory for AI coding agents
- Also built awesome-claude-code-toolkit, skillkit (universal CLI for AI agent skills), and k8sgames
Why it grew so fast
The “18% preparedness problem” is real and painful: most developers now use AI daily, but few can ship agentic pipelines professionally. This curriculum directly answers that frustration — and it’s free, MIT-licensed, and artifact-driven.
Community reception on Reddit was enthusiastic (“dope,” “Goat”) with healthy skepticism: senior engineers questioned how one person produces 500+ deep lessons without AI-generated slop, and early adopters hit empty folders and broken links in the initial Phase 0 rollout (since fixed). The fair assessment: it’s a remarkable structured map — go deep elsewhere (Karpathy for NN internals, papers for theory) where you need more than breadth.
Learning paths
| Your background | Start at | Time |
|---|---|---|
| Complete beginner | Phase 0 (Setup) | ~306h |
| Know Python, no ML | Phase 1 (Math) | ~270h |
| ML engineer, no deep learning | Phase 3 (Deep Learning Core) | ~200h |
| DL practitioner, no LLMs/agents | Phase 10 (LLMs from Scratch) | ~100h |
| Senior engineer, agents only | Phase 14 (Agent Engineering) | ~60h |
Not sure where you fit? Run the built-in find-your-level placement quiz (10 questions) and it drafts your path. There are also focused micro-paths: MCP (17 lessons) and Agent Skills (5 lessons).
Beyond the repo
The companion site (aiengineeringfromscratch.com) adds:
- 277-term AI glossary and interactive visual aids
- NLLB-200 machine translations — Spanish, Chinese, Hindi, Arabic, Turkish
- 6-volume book series in EPUB/PDF
- Claude Certification Academy: 33 dedicated lessons, timed mock exams, 295 original questions — prep for all 4 official Claude credentials (CCAO-F, CCDV-F, CCAR-F, CCAR-P), runnable directly via the
/claude-certificationcommand
Verdict
ai-engineering-from-scratch is the closest thing to a complete, free, self-serve AI engineering degree in 2026. The Build It/Use It method genuinely works — you can’t claim to understand attention if you’ve written it in NumPy first.
The honest caveats: no video lectures (repo-driven only), breadth over depth in places, and single-author scope limits. But for a structured map from linear algebra to agent swarms — with 511 artifacts you keep forever — it’s the best free deal in AI education right now.
Clone it, run the placement quiz, and start building. 🛠️