In an era where PyTorch and TensorFlow hide the mathematical guts of deep learning behind clean abstractions, this repo takes a refreshing step backward — in the best possible way. It implements a two-layer fully-connected neural network entirely from scratch using NumPy, with manually-derived backpropagation, and applies it to two classic benchmark datasets: CIFAR-10 and Fashion-MNIST.

What makes this project quietly compelling is the commitment to doing the math yourself. There's no loss.backward() magic here. The author writes out the forward pass, computes the softmax cross-entropy loss, then derives every gradient by hand and threads it back through the network. That's the exercise every deep learning course assigns, but few people bother to publish cleanly.

Why is that valuable? Because:

• Pedagogy: Reading a from-scratch implementation is often clearer than any textbook chapter on backprop. You see exactly how weight gradients relate to activations, and where the chain rule lands in code.
• Debuggability: When your fancy framework model won't train, understanding what a bare-metal implementation looks like helps you form better hypotheses.
• Two datasets, not one: Going beyond MNIST to CIFAR-10 forces the author (and reader) to confront that a two-layer MLP struggles on natural images — a valuable lesson about model capacity that motivates convolutional architectures.

Who would benefit? Undergraduate or self-taught ML learners who've watched Andrew Ng's course and want a compact reference implementation to study. Instructors looking for a clean starter template for their own assignments. And frankly, any working ML engineer who realized they've forgotten how softmax gradients actually work — a 30-minute read here will refresh the fundamentals faster than reopening a textbook.

It's a zero-star repo not because the work is poor, but because it's one of thousands of student projects — and that's precisely why it deserves a look before it disappears into the archive.