Most neural network tutorials jump quickly to frameworks and hide the math behind an API. This project goes the other way: a tiny two-layer network in NumPy, trained on XOR.
The forward pass turns inputs into activations. The backward pass uses the chain rule to move the error through the output and hidden layers, then adjusts weights. It is a small example, but it makes the mechanics of training visible in code. The point was learning how the pieces fit together, not building a general-purpose deep learning library.
Code: nn-from-scratch.