receiving, from a database system, a first data object comprising one or more data field values according to a data object schema of the database system, the first data object corresponding to customer data stored in a customer relationship management (CRM) system, and the one or more data field values comprising at least one non-numerical value, the at least one non-numerical value being a string value or a categorical value;

transforming, at the processing device, the one or more data field values of the first data object into one or more first features associated with a process flow, the transforming comprising generating at least one vector corresponding to the at least one non-numerical value;

routing, at the processing device, the first data object to a first path of a plurality of paths of the process flow using a random routing procedure, the routing comprising performing one or more operations associated with the first path using the one or more first features of the first data object;

training, at the processing device, a plurality of machine learning models based at least in part on an outcome of performing the one or more operations using the one or more first features of the first data object and a plurality of respective key performance indicators for the plurality of machine learning models;

inserting, at the processing device, a trained model of the plurality of machine learning models into the process flow at a decision point between the plurality of paths based at least in part on the training;

receiving, from the database system, a second data object comprising one or more second data field values corresponding to one or more second features associated with the process flow;

routing, at the processing device, the second data object to a second path of the plurality of paths for the process flow using the trained model inserted at the decision point and the one or more second features of the second data object; and

triggering, at the processing device, retraining of one or more models of the plurality of machine learning models based at least in part on one or more key performance indicators for the trained model inserted at the decision point failing to satisfy a performance threshold.