A good score on a test set proves less than it feels like
By the time you reach this course, you know how to train a model, engineer useful features, and evaluate it with the right metrics. That whole process happens in a controlled, static environment. You have a fixed training set, a fixed test set, and as much time as you want to iterate. Nothing is waiting on the result in real time, and the data isn't changing while you work.
Production is a different world. A real system needs the model to answer requests on demand, often within a strict latency budget measured in milliseconds, not minutes. It has to handle live traffic that looks nothing like a clean, pre-split test set, including malformed inputs, edge cases nobody labeled, and volume spikes. And it has to keep doing this correctly tomorrow, next month, and next year, even as the world the model was trained on keeps shifting underneath it.