GD AssistAI-Powered Claim Verification System
AI reads claim documents and pre-screens each claim against the plan, so examiners verify highlights, not whole files.
83%
Decisive and correct on 65 untouched holdout claims
- False AI rejections
- 0
- Cost per claim
- ~1.6 taka
- Need the AI arbiter
- 18%
- Read all documents
- 9 to 43 s
Checks on every claim
- Coverage and exclusions
- Identity, dates, and deadlines
- Amount reconciliation
- Annual limits
- Client
- GD Assist
- Location
- Dhaka, Bangladesh
- Industry
- Insurance, healthcare
- Engagement
- Custom build
- Timeline
- July 2026 to present
- Status
- Validated on real claims
About the client
GD Assist
Dhaka, Bangladesh
GD Assist, part of the Green Delta group, manages health insurance claims and cashless treatment in Bangladesh. We also built and run its claims platform.
Visit gdassist.comThe challenge
Examiners read every page of scanned, often handwritten claims in two languages.
GD Assist wanted software to do that first pass, without opening its claims database.
What we built
We designed, validated, and built the engine and a demonstration service.
- 01
Blind document extraction
A vision model reads every page without seeing the claim.
- 02
Deterministic checks engine
Compares what was read with the claim and plan, citing evidence.
- 03
AI arbitration
A text model decides what the rules refuse to, with a paste-ready remark.
- 04
Evaluation harness
Real adjudicated claims, with a frozen holdout, score every change.
- 05
Assessment playground
Switch plans and watch the same claim's verdict change.
Key features
Reads handwritten Bangla and English
Every flag quotes the text behind it.
Results
83%
Of 65 holdout claims decided correctly (54 of 65), with zero false AI rejections
Holdout run frozen at the proof-of-concept tag
36 of 36
Dental, optical, pregnancy and fertility, and policy-period rejections caught in development
100%
Development-set validation report
71 of 74
Approved development claims passed clean after tuning
96%
Development-set validation report
~1.6 taka
Cost per claim end to end, about 0.013 US dollars
Model benchmark and holdout cost ledger
18%
Of claims need the AI arbiter; rules alone decide the rest
18%
Development-set routing counts
How it works
Our role
Product framing, rejection analysis, prompt and extraction design, checks engine, arbitration, evaluation, demonstration service, and deployment.
Built with
- PyPython
- FaFastAPI
- vivision-capable large language model
- PiPillow
- DoDocker
- DoDocker Compose
Services behind this project
Advise & Run
AI Proof of Value
It began as exactly that: a measured test on real, already-decided claims.
Learn moreAdvise & Run
AI Readiness Assessment
Its scope came from analyzing more than eleven thousand historical rejections.
Learn moreBuild
AI Agents & Automation
The engine reads documents and pre-screens claims inside the examiner's workflow.
Learn moreWork with us
Test it on your own claims
Bring a handful of adjudicated claims; we will show what the engine reads, flags, and costs.