Forge exists because of a problem NumberGuesser ran into: a neural network trained from scratch on MNIST handled the original test set fine, then fell apart the moment it saw actual hand-drawn digits. Clean, centered, evenly-stroked training data doesn't prepare a model for how messy real handwriting actually is.

Forge is a macOS tool built to fix that. Describe the dataset you want in plain language: digit distribution, noise level, how many images; and it generates labeled digit images with realistic stroke variation (jitter, blur, grain), then splits everything into train and validation sets automatically.

It's built with SwiftUI on the front end, using Foundation Models to turn your description into config, and a Python engine underneath using PIL and NumPy. The two talk over an NDJSON stdio bridge, the same setup Rubrico uses.