micrograd-rs is a Rust implementation of Andrej Karpathy's micrograd: a tiny scalar-valued autograd engine with a small neural-network library on top of it, in the spirit of a PyTorch-like API. It's published on crates.io and built in CI with GitHub Actions.
I wrote it to understand how reverse-mode autodiff and backpropagation work at the level of individual scalar operations: each value remembers the operations that produced it, forming a computation graph that can be walked backwards to accumulate gradients. On top of that engine sits a minimal library of neurons, layers, and a multi-layer perceptron you can train with gradient descent. It was a personal exploration into neural networks, feeding into follow-on experiments like nn-sim, and is meant as a learning instrument rather than a production tool.
