I'm not using that in my own experiments since I don't want to worry about the performance of running a model on production, but seems worth a try.