It's a great model. We're embedding caffe as well. One thing I'd like to clarify though is that other neural network frameworks can also allow you to setup networks as well via yaml, json and the like.

Ours supports yaml/json. That's how we allow people to setup workers on spark hadoop (that then delegate numerical workloads to gpus or a lower level blas)

Pylearn2 also supports yaml.

Caffe has a lot of good ideas that are worth extending especially to the industry.

Our intent is for people to not touch java but pure bash.