[cercatrova]

We used TensorFlow Serving running in a docker container that is a rest endpoint for predictions. Then the backend can be agnostic to Serving. We personally used Node to query the container for whatever we needed.