MedShield AI AI-Powered Medical Insurance Fraud Investigation Platform

MedShield AI is a workflow-based multi-agent fraud investigation system built for Indian medical insurance companies to detect suspicious claims, reduce fraud losses, and accelerate investigation timelines.

Instead of acting like a chatbot, MedShield AI behaves like an autonomous insurance fraud investigator.

It uses specialized AI agents to:

extract structured claim information detect fraud patterns identify fraud rings challenge weak conclusions generate audit-ready final decisions

Built using Lyzr AI for multi-agent workflow orchestration, with support for GitHub backup, external integrations, and enterprise deployment.

Problem Statement

Medical insurance fraud causes massive financial losses across India.

Insurance companies like LIC, Star Health, HDFC Ergo, ICICI Lombard, and others spend weeks manually investigating suspicious claims involving:

phantom billing duplicate claims inflated treatment costs suspicious hospitals coordinated fraud rings repeated fraudulent doctors policy abuse by repeated patients

Manual fraud investigation is:

slow inconsistent expensive difficult to audit Goal

Reduce fraud investigation time from 2–3 weeks to under 30 seconds using autonomous AI agents and workflow-based investigation.

Built With Core Platform Lyzr AI (Architect + Workflow + Agent Studio) Additional Tools GitHub (backup + version control) VS Code (custom code + reliability) Perplexity API (optional external intelligence) PDF reporting system Database for persistent claim intelligence

Why Lyzr?

Lyzr helps rapidly build enterprise-grade AI agents and workflows without spending weeks on backend setup.

We use:

Architect

For:

system planning PRD generation workflow design agent creation architecture decisions Workflow

For:

chaining agents together defining execution order creating investigation pipelines Agent Studio

For:

refining prompts improving decision quality adding integrations tuning agent behavior Deployment

For:

live demo URL hackathon deployment investor-ready presentation

System Architecture Strict 5-Agent Workflow

Claim Input ↓ Extractor Agent ↓ Detective Agent ↓ Network Mapper Agent ↓ Adversary Agent ↓ Final Judge Agent ↓ Investigation Report + PDF + Dashboard Update

Agents Agent 1 — Extractor Agent Responsibilities Parse claim forms Parse invoices Parse prescriptions Parse discharge summaries Parse hospital reports Extracts patient name patient age policy number doctor name hospital name diagnosis ICD-10 diagnosis code medicines with quantity + unit prices procedures with billing amounts admission/discharge dates hospital stay duration claim total itemized bill total repeat claim history Output

Structured claim profile + extraction confidence score

Agent 2 — Detective Agent Responsibilities

Detect fraud patterns such as:

medicine priced 2x above standard market rate diagnosis vs treatment mismatch impossible procedures for diagnosis patient age vs treatment inconsistency inflated ICU charges duplicate procedure codes phantom billing upcoding total mismatch in billing repeated claims in short time windows Output fraud flags severity (HIGH / MEDIUM / LOW) exact reasoning initial fraud risk score (0–100)

Agent 3 — Network Mapper Agent Responsibilities

Check whether:

doctor hospital patient policy

are linked to suspicious previous claims.

Detect:

fraud rings suspicious provider repetition coordinated abuse patterns Output suspicious entity connections fraud ring detected (YES / NO) linked claims summary

Agent 4 — Adversary Agent (Core Differentiator) Responsibilities

Acts like a skeptical senior fraud investigator.

Its job is to:

challenge previous findings actively disagree if conclusions are weak identify missed fraud patterns force stronger investigation quality Rules never blindly agree if score is below 70, aggressively search for hidden risks revise fraud score if critical evidence is missed Output challenged findings contradictions found missed fraud signals revised fraud score

This is the biggest differentiator of MedShield AI.

Agent 5 — Final Judge Agent Responsibilities

Reads all previous outputs and produces:

Final Decision APPROVE MANUAL REVIEW REJECT Must Include final fraud risk score (0–100) confidence level top 3 reasons formal legal decision note recommended next action compliance-safe explanation audit-ready report

UI Pages Page 1 — Dashboard

Shows:

total claims analyzed fraud detection rate average fraud risk score high-risk claims today suspicious hospitals flagged suspicious doctors flagged last 5 claim verdicts Page 2 — New Investigation

Features:

claim upload form structured manual input Run Investigation button real-time live agent execution status

Example: Extractor → Detective → Network Mapper → Adversary → Final Judge

Page 3 — Investigation Report

Shows:

findings from all 5 agents fraud ring visualization investigation timeline large circular fraud risk gauge final verdict formal decision note downloadable PDF report

Investigation Timeline

Example:

10:42:01 — Extractor Agent parsed 23 claim fields 10:42:05 — Detective Agent found 4 fraud anomalies 10:42:08 — Network Mapper detected suspicious linked hospital 10:42:11 — Adversary Agent challenged low-risk assessment 10:42:14 — Final Judge issued REJECT decision

Impact Visual

Manual Investigation: 3 Weeks

vs

MedShield AI: 30 Seconds

This creates strong audit visibility and judge wow-factor.

Design Philosophy Enterprise Fraud Investigation War Room

Inspired by:

Bloomberg Terminal Investigation Command Centers Premium enterprise software Design Style dark professional UI red / amber / green severity indicators high trust interface data-dense layout minimal fluff enterprise licensing feel

The system should feel like software worth ₹50 lakhs—not a student chatbot.

Team Collaboration Strategy Team of 2 Person 1 — AI / Workflow Lead

Responsible for:

Architect Workflow design Agent Studio prompt optimization fraud logic testing GitHub backup Person 2 — Product / Demo Lead

Responsible for:

dashboard polish PDF report quality deployment demo story presentation final submission

Hackathon Demo Script Winning Line

“Every other system answers your questions. MedShield AI is the only one that questions your answers.”

Demo Flow Open Dashboard Paste suspicious claim Run Investigation Show live 5-agent workflow Show Investigation Timeline Show REJECT verdict Show fraud ring detection Show PDF report

This is the strongest judge moment.

Future Scope IRDAI blacklist integration hospital verification APIs insurer-to-insurer fraud intelligence sharing voice-based investigation support compliance automation multilingual investigator interface fraud trend forecasting Business Impact

MedShield AI can help insurers:

reduce fraud losses improve investigation speed improve audit reliability reduce manual workload improve fraud ring detection strengthen compliance reporting

This makes it not just a hackathon project, but a real insurtech startup opportunity.

Repository Structure medshield-ai/ │ ├── README.md ├── prompts.md ├── architecture.md ├── test-cases.md ├── demo-script.md ├── screenshots/ ├── reports/ └── deployment-notes.md Final Note

MedShield AI is designed to be:

hackathon-winning investor-demo ready enterprise believable scalable beyond prototype

This is not just another AI chatbot.

This is an autonomous fraud investigation system.