Health Insurance Appeals with Generative AI: From Lab to Production
summary
A system generates patient appeals from denials by fine‑tuning open‑license LLMs on California IMR records and synthetic examples, using OxaloTLE-style tooling, Hugging Face checkpoints and Lambda Labs GPUs; fine‑tuning costs about $112 fine-tune and they iterate to version 0.5. Serving runs on consumer hardware with RTX 3090 serving or Kubernetes, a Django frontend does client‑side redaction, and real users report some appeals win while others still fail, so it’s useful but imperfect.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:25Content warning
- 3:38What's the problem?
- 6:13How to use computers to fix the problem?
- 7:13What's needed to tune a LLM?
- 10:36Putting it together
- 11:38Downsides
- 13:04What models did we make?
- 14:21Model fine tune config
- 15:40Model serving
- 18:38Frontend?
- 19:46Demo
- 25:18How to do better for the next model?
- 27:55Links
- 28:40Outro
related talks
31mAI For Good: Fighting Health Insurance with AI
10mAI-Powered Dental Claims Processing
41mAn Art, a Science, or Your Next Job Title?
29mGenerative-AI: Usecase-Specific Evaluation of LLM-powered Applications
22mHarnessing AI for Underwriting in Insurance
18mGenerative AI in Financial Industry
26mWelcome to the AI Jungle! Now What?
50mWhy use AI in insurance underwriting and claims? (PRT271)