Built a fully functional convolutional neural network (CNN) from first principles using NumPy, without relying on any machine learning libraries. Implemented a custom automatic differentiation (backpropagation) engine, convolution layers, ReLU activations, fully connected layers, and optimization algorithms (SGD and Adam), forming a complete end-to-end training pipeline.

Trained and evaluated the model on the MNIST dataset and rigorously benchmarked its performance against a matched PyTorch baseline under identical architectures, initializations, and hyperparameters. Verified numerical correctness through close agreement in training loss, test accuracy, gradient norms, and optimization dynamics.