But does it really speed up numerical libraries like numpy and pandas? I thought it only works on pure python code.