How hard would it be to utilize Neanderthal for your example in the article?
The new eager execution in TensorFlow is supposed to integrate well with Python debugging and immediate run-time errors. Is it possible to get better debugging information from Clojure?