How we found a simple way to reduce the memory consumption of production models by up to 80% with no loss in predictive power!
