Totally. And I think it’s a solid strategy. However, there’s certainly an interest (internally and externally) in providing a better inter-operability story between the two. I imagine something like using Keras for model creation (and possibly training) and running (either on mobile or the cloud) on some Caffe 2 deployment.