It is a bittersweet change, in that there was a lot of hard work in the old code that has proven to be no match for TensorFlow (believe it or not, one EE grad student's 4 months of code was no match for Google's team of engineers' years of code :-)
Ripping everything out and starting anew was not taken lightly - and only done when it was unequivocally certain that the raw low-level control of the weights and training/evaluation loops could be maintained, while reaping the benefits of the optimized functions that TensorFlow offers.
It will likely be challenging to convey the purpose and usefulness of LowPy to those who need it, but it will be worthwhile. There are many developments to be made, and lots of incoming commits to look forward to!