Draw a digit and see it classified live, with inference running on the GPU.
MNISTify makes digit recognition something you can play with instead of read about. You draw a number on a canvas and the prediction updates as you draw, with a live chart of what the network is actually outputting so you can watch the confidence move rather than just read a label.
Two architectures are available and can be compared side by side: a multilayer perceptron, which is fast and light, and a convolutional network, which is more accurate. You can inspect the model's input and compare inference time between them. Training runs on the MNIST dataset through TinyGrad; inference then runs in the browser on WebGPU, so the whole thing costs nothing to host and stays responsive while you draw.
