README.md

metadata

title: ClaimReady — Claim Submission Check
emoji: 🏥
colorFrom: blue
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
short_description: Catch claim errors before you submit — open model
tags:
  - build-small-hackathon
  - document-ai
  - vision-language
  - multilingual
  - healthcare
  - track:backyard
  - achievement:offgrid
  - achievement:offbrand

🏥 ClaimReady

An assistive pre-check for hospital health-insurance claims — built for hospital administration staff. ClaimReady helps hospitals identify potential compliance issues before submitting claim documents to insurance providers. It analyses uploaded documents against package-specific treatment guidelines and provides actionable feedback on missing, incomplete, or non-compliant documents, allowing hospitals to correct issues early in the process.

It runs entirely inside the Space on a small open model — Gemma 3 12B (≤ 32B) — with no cloud inference API.

🎥 Demo & Links

• ▶️ Demo video: https://youtu.be/A4PCHp89bsQ
• 📣 Social post: https://www.linkedin.com/posts/ullivinaybabu_build-huggingface-gradio-share-7472141735138119681-80Th/
• 💻 Source code (GitHub): https://github.com/vinayulli/claim_ready
• 🚀 Live app: this Space — open the "Load a sample claim" buttons at the top to try it in one tap.

📋 Overview

ClaimReady is used by hospital administration / billing staff. For this prototype it uses the PMJAY (Pradhan Mantri Jan Arogya Yojana) scheme as an example, where hospitals must follow Standard Treatment Guidelines (STGs) when submitting claims. Because these guidelines vary across treatment packages, the application evaluates each claim against the selected package's requirements.

The goal is simple: reduce claim rejections, minimise processing delays, and improve operational efficiency for hospitals and claim teams.

🚩 Problem Statement

Health-insurance claims — for example under India's Ayushman Bharat / PMJAY scheme — require a specific set of supporting documents that must satisfy the applicable clinical / treatment guidelines for each procedure and stage (pre-authorization / claim).

• Every claim must include all mandatory documents and meet defined content conditions for the procedure and stage.
• A missing document — or a value that doesn't meet a condition — can lead to claim rejection, delays and rework.

✅ Solution — What ClaimReady Offers

• 📄 Reads every uploaded document with on-device OCR — images and PDFs.
• ✅ Verifies the set against the required-document checklist for the selected package and stage.
• 🔎 Evaluates content rules (thresholds, conditions) against the values it actually reads.
• 🌐 Handles mixed-language documents — e.g. English + Hindi / Telugu in the same record.
• 🖼️ Built-in document viewer — preview every page (images and PDF pages) before checking.
• ⚠️ Surfaces missing, incomplete, or non-compliant documents early, with supporting evidence — as an assistive pre-check.

🔧 How It Works

1. Select a package + stage → the app loads the required-document checklist and content rules (data-driven from packages.json).
2. Upload documents (or click a sample) → PDFs and images are rendered to page images (PyMuPDF) and shown in the viewer.
3. Run the check → the document images + the checklist + a strict review prompt go to Gemma 3 12B, which OCRs and reasons in a single pass and returns structured JSON.
4. Review → rendered as a clear ✅ / ❌ / ⚠️ assistive review with evidence and an action list. (Low-confidence "present" documents are flagged as gaps, to stay conservative.)

📈 Scope & Scalability

• Currently supports four treatment packages as a proof of concept.
• The logic is data-driven (packages.json), so the architecture is scalable and can easily be extended to support many more packages, insurance providers, and compliance frameworks.

🧱 Built Small

Model	google/gemma-3-12b-it — a small, open ≤ 32B model
Runtime	Hugging Face Transformers on ZeroGPU — runs inside the Space, no external / cloud inference API
Stack	Gradio · PyMuPDF · Pillow · 🤗 Transformers

🩺 Note

ClaimReady is a decision-support tool — it highlights likely gaps for review and does not approve or reject claims. The sample claims in the app are synthetic (fictional patients) — no real patient data.

---

Built by @vinaybabu for the HuggingFace Build Small Hackathon · Backyard AI track.