numba/numba ? reverse-engineered prompt

Build me a Python package like Numba that speeds up numerical code by compiling selected Python functions just in time into machine code.

I want it to work best with NumPy style code, simple loops, math, and array operations, and to handle a useful subset of numeric Python rather than every possible Python feature. It should also support parallel loop execution where possible, plus a path for GPU accelerated code and the ability to create ufuncs and C callbacks.

Please make the project feel like a real open source library, with a clean install, basic docs, tests, and a simple demo showing how someone would decorate a function and get faster execution. If you need to check current best practices or docs while building it, go ahead and look them up online.