The thing to focus on for rai is LLVM. I'd like to use rai for data-intensive modeling as well. Something that needs horsepower on the host cpu. Most of these things do not need "structural" flexibility that racket offers. The result is state machines performing math ops. Let's find something that does that.