Holy shit, this looks amazing!

I see you've got examples for NLP use cases in your docs. Can't wait to read them. Embeddings are a constant source of complexity when I'm trying to move certain operations to Lambda, this looks like it would speed the initializations up big time.