At Chatty Insights, I built the moderator agent and runtime for an AI-moderated research platform, including session state that can resume after a crash and context for dynamic probing. It served about 170 live interviews in one minute at peak.
I also benchmarked model decisions against production traffic and incumbent-model noise, and held back a check that wrongly flagged short but complete answers. As an independent consultant, I built a LoCoMo-based agent memory system that raised long-horizon recall from 15% to 75% on the benchmark, and published a hard-case RAG benchmark.
