After building a Keras-like library I got obsessed with figuring out how autograd works.
Accordingly, here's my toy attempt to build a PyTorch-like library (again, using NumPy).
As usual, there is a corresponding series of livecoding videos on my YouTube.