Whenever you're ready, walk me through how you'd approach this.
How would you design a system to serve LLM responses at scale with low latency, cost controls, and safe fallback behavior for The Boston Consulting Group client work? Discuss how you would define service-level objectives, evaluate quality before launch, route requests across models, and handle overload, provider failures, hallucinations, and prompt injection. Deliverables: 1. Propose the serving architecture and request flow. 2. Explain model routing, caching, batching, and cost controls. 3. Define offline and online evaluation. 4. Describe fallback and safety mechanisms.