Whenever you're ready, walk me through how you'd approach this.
How do you evaluate the cost-to-performance trade-offs when deploying large language models versus smaller task-specific machine learning models?
Discuss the decision framework you would use for a production ML system. Cover model quality, latency, throughput, infrastructure cost, operational complexity, data and retraining requirements, observability, and failure handling. Explain when you would choose an LLM, a specialized model, or a routed combination of both, and how you would validate the choice offline and online.