LLM output quality monitoring

Design the production quality-monitoring system for a customer-support LLM assistant handling millions of conversations a day. Unlike a classifier there's no single accuracy number, and outputs are free-form text that can be subtly wrong, unhelpful, off-policy, or hallucinated. You need to detect quality regressions (e.g. after a prompt change or model upgrade) quickly, flag harmful/hallucinated responses, and do it without humans reading every conversation. Design the monitoring + evaluation system for generative output at scale.