thekiku/neural-network-from-scratch ? reverse-engineered prompt

Build me a Jupyter notebook project that teaches and demonstrates a neural network built completely from scratch with NumPy, without using TensorFlow or PyTorch.

I want it to walk through a simple feedforward model with dense layers, forward and backward pass, ReLU, sigmoid, linear, and softmax activations, plus common losses like categorical cross entropy, binary cross entropy, MSE, and MAE. It should train on the spiral dataset, show accuracy and loss while training, and include examples for classification and regression. Please also include regularization, dropout, and a few optimizers like SGD, momentum, AdaGrad, RMSProp, and Adam, along with learning rate decay.

Make it easy to run in Jupyter, use clear notebook cells, and include plots so I can see the data and training results. If needed, look up current docs online for any library details. Also add saving and loading the model and a simple prediction example on new data.