How would you approach the transition from using a model for batch predictions to real-time predictions?