From Static to Continuous: A Design Pattern for Real-Time ML Systems
There's a structural choice in ML system design that gets made early, usually implicitly, and then dictates the entire shape of the resulting product. The choice is: does the model's state update between user interactions, or in batch after the session ends?