To be honest it was straight forward. We used BERT-base (although maybe albert would be better choice now). We fine tuned on our corpus (took ~1 month on a V100, if you do it in cloud with multiple GPUs would be faster, but we were just playing around/dumped it on one or our ml dev servers). Finally we do inference using a containerized BERT serving https://pypi.org/project/bert-serving-server/ this output is consumed by some other elements of ml pipeline.