Context:

There are two RevertRisk models. The Multilingual (RRML) is better for anonymous users, but it has some performance issues. The Language Agnostic (RRLA) one is faster. The recommendation now is: use the RRLA for registered users, and use the RRML for IP edits.

An option would be to create a "meta-model" that decides which model to use, transparently to the end-user. The advengate of this solution is that it would allow us to update to the best model everytime that we do a modification (i.e. improve one of the models). The cons of this solution could be that it might add some extra-latency.

ML team needs to verify how to do it. A possible solution is using Kserve's Inference Graph
https://kserve.github.io/website/0.9/modelserving/inference_graph/#concepts