Various embodiments of the present disclosure provide a machine learning request handling technique that improves the functionality of a computer in various aspects. The techniques include receiving a model ensemble prompt request comprising a user identifier and a model prompt. The techniques include generating, based on the user identifier of the model ensemble prompt request, a modular model ensemble that references a plurality of machine learning models. The techniques include generating, based on the model prompt, a plurality of asynchronous model prompts for the plurality of machine learning models, respectively. The techniques include inputting the plurality of asynchronous model prompts to the plurality of machine learning models, respectively, to generate a plurality of candidate model outputs and outputting a model ensemble prompt response that includes a selected output from the plurality of candidate model outputs.