I build the intelligence layer between industrial systems and the people who run them.

Move across the network to drive it — your pointer is the input vector. Five inputs, two dense hidden layers, a softmax output. It is really training as you watch: forward pass, cross entropy against a rotating target class, gradients back through the same layers, one weight update per cycle. The curve above each hidden layer is its activation function, plotted from its own definition; it swaps every few cycles and the weights start over, because they were fitted under the old one.