[goatinaboat]

But for machine learning or numerical algorithms, I found it quite hard to make k nearly as fast as vanilla NumPy code

Observe that KX’s approach to ML is... to just wrap Python https://code.kx.com/v2/ml/