HudHatman/impulse-ml — reverse-engineered prompt

Build me a small Node.js machine learning library that lets me create and train simple feed forward neural networks from JavaScript. I want to define a network by stacking layers like sigmoid, relu, tanh, and softmax, then train it with choices like gradient descent, momentum, adagrad, rmsprop, or adam. It should support mean squared error and cross entropy, batch training, learning rate, regularization, batch size, and iteration settings.

Please make it able to load datasets from CSV files, do basic cleanup like shuffling, min max scaling, and handling missing values, then let me save a trained model to JSON and load it back later. I also want a native C++ addon for the heavy math so it runs in parallel on the CPU with OpenMP, while still feeling easy to use from Node.

Include a couple of working example scripts, like Iris and MNIST style classification, plus simple docs and build instructions. If anything is unclear, look up current Node addon docs and use sensible defaults.