This is my neural network implementation, written entirely from scratch. It uses a simple feedforward neural network with two dense layers: the first containing 512 neurons and the second containing 256 neurons. The network was trained using my C# implementation of a neural network with CUDA support, which I also ported to TypeScript. The weights and biases are stored in a file on the server and loaded into the model running on your device. When you draw, the model predicts 10 values, each representing a number. The higher the value of a specific neuron, the more likely it is that the number you have drawn corresponds to that neuron. The model was trained on the MNIST database.