jnhl

You could always go deeper and from some points of view, it's not "from the ground up" enough unless you build your own autograd and tensors from plain numpy arrays.

Untraining models is an open research question and not par for the course like you're implying