A deep learning framework built from scratch.

Automatic differentiation, neural networks, and training, from first principles, one number at a time. Understand how PyTorch actually works by building the engine underneath it.

Start the course View on GitHub

Learn, the course

23 lessons across 6 units, with interactive demos. Derivatives, the computation graph, backpropagation, neural networks, and training, the ideas you need to build the framework yourself.

Start learning

Docs, install & use

Installation, the API, and usage examples for the framework itself. Arriving with the first release.

Read the docs

What makes it different

From first principles

No black boxes. The autograd engine starts on single numbers so every operation and every gradient is visible, then climbs to tensors.

Taught, not just shipped

A full interactive course ships alongside the code. It teaches the concepts; you write the framework, which is where understanding sticks.

A real path to a real framework

v1 is a scalar autograd engine that trains a net. v2 climbs to tensors, then lazy evaluation and kernel fusion, the parts that make a framework fast.