How would you build a real-time monitoring and observability pipeline for deployed LLM agents to track drift and latency at Ema?
