I don't know if that's the recommended way, but what I sometimes do is train a bento service together with the experiments, and log its name with MLFlow, so when I decide on the best model I already have a service waiting to be deployed.