Trust, but Verify: Lessons from Deploying LLMs in a Large Health System
summary
Stanford Health Care deploys an LLM chatbot called Chatter (ChatEHR) to generate real-time chart abstractions and referrals, routes clinicians to multiple models, and logs clinician agree/disagree decisions to measure human-in-the-loop accuracy. A custom Python evaluation harness, MEDHELM benchmark suite, scores task-specific performance whenever external models update. Automated Quarto reports run fact decomposition hallucination rate checks; one legacy model shows a markedly higher hallucination rate, prompting deprecation and model switching.
AI-written summary · about
chapters (8)
- 0:00LLMs in healthcare intro
- 3:13LLM reliability challenge
- 3:56MLOps for LLMs
- 6:48Human-in-the-loop monitoring
- 10:04Handling model updates
- 13:06Monitoring failure modes
- 16:39Overcoming LLM intimidation
- 19:38Q&A
related talks
13mNew Challenges in Software Quality Brought by GenAI and LLMs
33mLessons Learned Applying Large Language Models in Healthcare
48mUnlocking LLM Potential: From Exploration to Integration
17mArchitecting LLM-Based Chatbots using free open-source LLMs
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
47mCan You Trust Your ChatBot: Techniques for Testing LLM Responses
26mReducing Hallucinations and Evaluating LLMs for Production
38mUnraveling Multimodality with Large Language Models