I trained this on my own but for my CS 189 intro to ML class at Berkeley. The goal here was to take all of my whiteboard math for deriving backpropigation and train something with it.

Model arch is standard MLP ReLU activation (and softmax on output) with layers 784, 128, 64, 10 and 109,386 params.

You can simply draw on the square below and see what the model does in real time. You can also download the model by just looking at the web source.