Whenever you're ready, walk me through how you'd approach this.
How would you design a real-time inference pipeline for a high-traffic application at a company?
Describe how you would clarify requirements, size the system, choose model-serving and feature-serving components, and handle online inference at scale. Cover latency, availability, deployment, evaluation, monitoring, feature drift, training-serving skew, and graceful degradation.