AI Engineering from Scratch<br>AI">

FIG_000 &middot; curriculum v1.0 · 2026<br>open source · MIT

AI Engineering<br>from Scratch

416 lessons. 20 phases. Every algorithm built from raw math before a single framework gets imported.

Maintained by Rohit Ghumare and contributors. Run on your own machine.

FIG_006 · the stack<br>memory · reasoning · kb

Three repos compose into the agent stack the curriculum teaches.

REPO · REASONING<br>agentbrain<br>★ ...

agent -->

STEERS

AGENT LOOP

function · worker · trigger<br>+ queue · persistence · authz<br>FROM THE CURRICULUM

memory -->

READ / WRITE

REPO · MEMORY<br>agentmemory<br>★ ...

akbp -->

ANSWERS

REPO · KNOWLEDGE<br>akbp<br>★ ...

SAME AUTHOR · OPEN SOURCE<br>FULL STACK BELOW · SCROLL TO FIG_006

How this works

Most AI material teaches in scattered pieces. A paper here, a fine-tuning post there, a flashy agent demo somewhere else. The pieces rarely line up. You ship a chatbot but can't explain its loss curve. You hook a function to an agent but can't say what attention does inside the model that's calling it.

This curriculum is the spine. 20 phases, 416 lessons, four languages: Python, TypeScript, Rust, Julia. Linear algebra at one end, autonomous swarms at the other. Every algorithm gets built from raw math first. Backprop. Tokenizer. Attention. Agent loop. By the time PyTorch shows up, you already know what it's doing under the hood.

Each lesson runs the same loop: read the problem, derive the math, write the code, run the test, keep the artifact. No five-minute videos, no copy-paste deploys, no hand-holding. Free, open source, and built to run on your own laptop.

Current Progress

Finished Lessons<br>bar<br>0 / 0

Phases<br>bar<br>0 / 0

Languages<br>bar

Glossary Terms<br>bar

Curriculum · 20 phases · 416 lessons

Tap a phase to expand its lessons. Each one ships when its math, code, and test are all written.

Complete<br>In progress<br>Planned

Progress saved in browser only<br>Reset progress

From the same author

Memory + reasoning + knowledge protocol

Three open-source repositories that compose into a full agent stack. The curriculum teaches the primitives; these tools ship them in production.

FIG_006 · the agent stack<br>memory · reasoning · knowledge

Three repositories — agentmemory, agentbrain, akbp — composing into the agent stack the curriculum teaches.

AGENT LOOP

function · worker · trigger<br>you build this in Phase 14<br>FROM THE CURRICULUM

MEMORY SURFACE<br>agentmemory<br>persistent memory for AI agents.<br>★ ...

REASONING SURFACE<br>agentbrain<br>evidence-first reasoning loop.<br>★ ...

KNOWLEDGE SURFACE<br>akbp<br>agent knowledge base protocol.<br>★ ...

AGENT (one label, two arrowheads) -->

READ / WRITE STATE

STEERS

AGENT (one label, two arrowheads) -->

ANSWERS QUERIES

EIGHT PRIMITIVES<br>function · worker · trigger · runtime · HTTP/RPC · queue · session persistence · authorization policy<br>Phase 14 teaches the primitives. The three repos above are the same primitives, shipped.

Three repositories, one stack: agentmemory writes durable state, agentbrain steers the loop from evidence, akbp answers knowledge queries between sessions and across agents.

agentmemory

Persistent memory for AI coding agents. The state surface from Phase 14, productionized.

View on GitHub →

agentbrain

Evidence-first operating system for agents. The reasoning + verification surfaces, wired end-to-end.

View on GitHub →

akbp

Agent Knowledge Base Protocol. The handoff + knowledge layer between sessions and across agents.

View on GitHub →

Colophon

The entire curriculum is on GitHub. Clone it, fork it, learn at your own pace. No paywall, no signup. Every lesson has runnable code in Python, TypeScript, Rust, or Julia, depending on what fits the concept best.

git clone https://github.com/rohitg00/ai-engineering-from-scratch.git<br>cp