{
  "id": 15299047,
  "title": "ClawbackMD: local Gemma drafts for AI-assisted insurance denials",
  "url": "https://urgent.news/2026/10/17/clawbackmd-local-gemma-drafts-for-ai-assisted-insurance-denials",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-17T23:07:49.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/walsoup/clawbackmd-local-gemma-drafts-for-ai-assisted-insurance-denials-1n3g"
  },
  "original_language": "en",
  "account": "ClawbackMD is a local open-source AI tool designed to draft appeal packets for insurance denials in a way that mimics AI-generated text. The software takes a denied claim, analyzes the synthetic chart and payer policy, and produces a verdict, evidence, letter, and talking points. The default model is in-memory, meaning it doesn't require Docker. If Ollama/Gemma is unavailable, cached human-reviewed drafts are still functional. Every letter generated is a draft, not an e-signed document.\n\nThe primary audience for ClawbackMD are hospital billing and revenue cycle management (RCM) teams who often spend hours on AI-assisted denial processes. The tool is designed to run on a local laptop without exposing protected health information (PHI) to a hosted API. A demo repository is available at https://github.com/walsoup/clawbackmd, and a live demo can be accessed at https://clawbackmd.vercel.app.\n\nTo run the demo, users need to install the necessary packages, create a `.env.local` file from `.env.example`, and start the application using `npm run dev`. Users can opt for the optional model, Ollama Gemma3, and serve it with `ollama serve`. The tool supports five synthetic claims upon initial launch. To generate a draft, users can input CLM-001 (appeal) or CLM-005 (documentation gap) and query the attending.\n\nThe tool relies on open-weight Gemma via any OpenAI-compatible endpoint, typically local Ollama. It includes built-in guardrails to ensure that evidence and policy quotes are verbatim substrings of the source note or policy. Sentences in the letter that contain clinical information must appear in the validated evidence or claim facts; otherwise, those sentences are removed. The model can only cite payer policy codes, and any server-side references to CFR/statutes are pulled from the plan_type, not invented by the model. Upholding the denial status means querying the attending for documentation gaps instead of automatically issuing a write-off.\n\nThe user interface is built using the Linen design system, featuring a linen surface, forest primary color, and hairline strokes. Open innovation is a key aspect of ClawbackMD, as PHI remains local and never leaves the machine when running Gemma on Ollama. The tool eliminates the need for per-token billing, as billing staff can run the drafting loop all day without incurring additional costs. The model is swappable, allowing users to point the LLM_BASE_URL or LLM_MODEL to another OpenAI-compatible endpoint without rewriting the app. The system is designed to \"fail closed\" on quotes, meaning if the model invents a finding, validators will discard it rather than producing a confident fake. However, this tradeoff necessitates cached drafts and a quote scrubber to ensure the demo does not inadvertently invent clinical facts.",
  "summary": "This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 2 — Side Hustle What I Built ClawbackMD drafts appeal packets for insurance denials that look AI-written on the payer side. You pick a denied claim, it reads the synthetic chart + payer policy, and it returns a verdict, quote-backed evidence, an editable letter, and peer-to-peer talking points. The model is optional.…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}