AI For Good: Fighting Health Insurance with AI
summary
A nonprofit pipeline turns published state independent medical review decisions into training pairs, fine-tunes Gemma/Mistral-derived models, and serves them on a rack using vLLM and Ray with client-side PHI handling and regex filters as fallbacks; fine-tunes cost about $112 fine-tune cost per run and the project reports a few thousand generated appeals so far, plus live outages during production rollouts and occasional demo failures, but overall the system produces usable appeal letters and a frontend that real users can submit.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 5:23Why health insurance is broken
- 8:53Gathering training data for AI
- 13:09Building the AI model
- 19:11Fine-tuning the model
- 23:07From lab to production
- 24:31Demo and project status
- 29:27Q&A and closing
related talks
29mHealth Insurance Appeals with Generative AI: From Lab to Production
18mAI That Pays: Lessons from Revenue Cycle
10mAI-Powered Dental Claims Processing
17mAI Innovation in Health
25mAI for healthcare: Scaling Access and Quality of Care for Everyone
16mAI-Driven Claims Adjudication: Healthcare Efficiency & Fraud Prevention
59mArtificial intelligence in Healthcare
16mGenAI in Healthcare Claims: Processing and Detecting Fraud