Describe a time your model performed well offline but degraded in production. How did you architect a solution to catch and fix this?
