TL;DR
micrograd is a tiny engine implementing backpropagation (automatic differentiation) for scalar values and a PyTorch-style neural network library, designed for educational use.
Key features
Autograd engine: Builds a dynamic DAG to perform backpropagation on scalar values, implemented in about 100 lines of code.
Neural network library: Includes neurons, layers, MLP, etc., composed of about 50 lines of code.
PyTorch-like API: Uses Value objects to define operations and backward() to compute gradients.
Visualization support: Can visualize computation graphs and gradients via graphviz.
When to use it
When learning how automatic differentiation and backpropagation work.
For simple neural network implementation practice or educational materials.
Suitable for conceptual understanding and experimentation rather than production use.