Whenever you're ready, walk me through how you'd approach this.
Design an end-to-end system that ingests thousands of claim photos per minute, runs them through an ensemble of ML models, and returns a repair estimate in under two seconds for GEICO.
Provide a practical architecture, implementation approach, and deployment plan using specific technologies. Show how you handle ingestion, preprocessing, parallel model inference, aggregation, response serving, retries, idempotency, malformed images, model versioning, and observability. Include representative code or configuration, a data-flow diagram, and tests that demonstrate the latency target and failure behavior.