how to include post-processing step in ensemble model ?

[Aktharkazi34]

I am building and deploying a recommendation model using the NVIDIA Merlin framework. The model has been successfully trained and deployed to the Triton Inference Server as an ensemble model. Currently, when generating recommendations via HTTP requests, I need to perform several post-processing steps. These include converting logit values to probabilities, mapping transformed recommendation IDs back to their original values, and excluding certain items from the recommendation list based on specific business rules.

Is there a way to integrate all of these post-processing steps within the ensemble model as a final step, so that each request directly outputs refined recommendations?

Here is my current model repository structure.