Design a multi-model serving platform for LLMs with autoscaling and cold-start handling under a cost ceiling.

The tell that someone has actually run LLM serving is that they scale on token throughput and KV-cache headroom, not requests, because one request can cost a hundred times another. The held-back follow-up is 'a cold model just got a request, what does the user see, and what's your number?', and a serious answer names tens of seconds for a cold large model and a tiered-residency plan to avoid it on hot paths. Watch for the candidate who reaches for scale-to-zero everywhere because it demos well; on a user-facing route that's a self-inflicted brownout, and saying so is the senior move.