Whenever you're ready, walk me through how you'd approach this.
How would you design a feedback loop to continuously improve model performance based on user interactions at a bank?
Discuss how interactions become reliable training data, how models are retrained and deployed, and how you would measure improvement without creating harmful feedback loops. Cover online and batch components, monitoring, experimentation, privacy, and rollback. State any scale assumptions you need rather than treating them as known bank facts.