Racket is an amazing language for prototyping ideas that you don't understand yet.

At $dayjob I'm using it to test what novel geometries of deep learning models would look like. Being able to redefine any part of the stack for any reason is a superpower you don't know you need until you do.

A great place to start is the little learner which holds your hand until you get opinionated about what the underlying primitives should look like. E.g. what if we used sparse tensor representation?

You might like having a go at Lush. It has fallen out of favor of late but is a very interesting language/system.

That sounds kind of amazing. But you're not actually doing the machine learning in Racket, are you? Is your Racket code generating other code like PyTorch?