This was a useful read. I ended up checking out the GeekyAnts article you referenced, and one thing both pieces reinforce is that successful AI products depend as much on solid engineering and operational planning as they do on the models themselves.