It's been done: https://github.com/pola-rs/polars

But I'm sure there's always room for improvement

It’s not like polars is a drop-in replacement, it has a totally different API.

You wrote "it has a totally different API", did you mean "it has an actually sane API?" Because that's what I think of when I compare pandas to polars.

Not if you've legacy apps using pandas everywhere.

Same API means I can import polars as pd and be done with it.

This is a curious reply for me. I would think that there are very few parts in pandas that could be sped-up by reimplementing them with a compiled language. Pandas is plenty fast for the built-in methods, it only gets slow when you start interfacing with Python, e.g. by doing an `.apply` with your custom Python method. Obviously this interfacing part is impossible to speed up by reimplementing parts of pandas (you'd need a different API instead).

I remember, when trying to squeeze some performance out of it, that a lot of the overhead came from it trying to infer types.

would give https://github.com/modin-project/modin a shot

The answer would then be to have a look at polars.