Saved Keras model meta data and weights, wrote Racket code to read weights, use low level matrix library, relu , sigmoid, to use trained models. So, just use TensorFlow for training.

This works, but the performance was not great. I want something more robust and performant.