Whenever you're ready, walk me through how you'd approach this.
How would you approach the architecture of an AI service that personalizes user experiences on American Express based on historical data? Discuss how you would clarify requirements, separate offline training from online inference, and design the retrieval, ranking, and re-ranking stages. Explain your approach to data freshness, privacy, evaluation, feature drift, training-serving skew, monitoring, and graceful degradation. State which scale and latency assumptions you would request before finalizing the design.