There used to be Swift for Tensorflow but google killed it and then Chris Lattner created Mojo. AFAIR the punchline was that Swift was either too complicated for data scientists or they just preferred to use python that already knew.
Seift has great C, C++ and Rust interop.
And despite Swift for Tensorflow being killed, there’s still decent python interop.
Now, setting up a project for that is probably quite hard because the documentation is non-existent.
I would happily have something 50-100% slower than C++/Rust with good interop, alas there seems to be very little / nothing. Cython for various reasons isn't ideal.