APL wasn't originally a programming language at all. Ken Iverson designed it as a mathematical notation to express and reason about computer algorithms. "Notation as a tool of thought," is how he thought about it. It was implemented as a programming language years after he created the notation in 1960.

An APL-like Haskell DSL could be interesting, but to match APL's expressiveness you'd basically need to reimplement all of APL. For maximum generality, one could just skip both APL and Haskell and just use the lambda calculus. I found that a bit hard to work with, though.

As far as Numpy and all, they are all directly descended from APL. The difference is that it takes 10 lines of Python to match 10 characters of APL. While the array languages' terseness can be excessive, doing it in Python is not very pleasant either.

Anyway, here's somebody who knows more than I do talking about what they like about APL's successor, J: https://www.youtube.com/watch?v=RWYkx6-L04Q