Solarys Fraud Detection Demo

A demonstration of AI-powered healthcare fraud detection using multi-vendor AI agents that communicate via open protocols. This project showcases how Claude (Anthropic) and Gemini (Google) work together through the A2A (Agent-to-Agent) protocol, while accessing fraud detection tools through MCP (Model Context Protocol).

Engineered by the Solarys Group as a proof-of-concept for government healthcare fraud investigation systems.

Try the Live Demo →

Table of Contents

• Overview
• Architecture
• Protocol Deep Dive
  • MCP (Model Context Protocol)
  • A2A (Agent-to-Agent Protocol)
• Key Components
• Fraud Detection Tools
• Data
• Project Structure

Overview

Two AI agents collaborate on fraud investigations:

1. Detection Agent (Claude) - Primary orchestrator, runs initial analysis, delegates when needed
2. Investigation Agent (Gemini) - Deep-dive specialist for complex cases

Communication happens over open protocols:

• MCP - Tool access (queries, anomaly detection, pattern matching)
• A2A - Agent-to-agent delegation

Architecture

flowchart TB
    subgraph Frontend["Frontend (Next.js)"]
        UI[Investigation UI]
    end

    subgraph API["API Server (Express)"]
        Routes[SSE Streaming]
    end

    subgraph Agents["AI Agents"]
        Claude["Detection Agent<br/>(Claude)"]
        Gemini["Investigation Agent<br/>(Gemini)"]
    end

    subgraph MCP["MCP Server"]
        Tools["9 Fraud Detection Tools"]
    end

    subgraph DB["PostgreSQL"]
        Data["558k claims / 5.4k providers"]
    end

    UI -->|"HTTP/SSE"| Routes
    Routes -->|"A2A Protocol"| Claude
    Claude <-->|"A2A Protocol"| Gemini
    Claude -->|"MCP Protocol"| Tools
    Gemini -->|"MCP Protocol"| Tools
    Tools -->|"Prisma"| Data

Protocol Deep Dive

MCP (Model Context Protocol)

MCP is an open protocol by Anthropic that standardizes how AI models interact with external tools. Both agents in this demo access fraud detection tools through the same MCP server.

Request flow:

AI Agent                    MCP Server                    Database
   │                            │                            │
   │ ──── ListTools ─────────▶  │                            │
   │ ◀─── [9 available tools] ─ │                            │
   │                            │                            │
   │ ──── CallTool ───────────▶ │                            │
   │      "find_anomalies"      │ ──── SQL Query ─────────▶  │
   │      {providerId: "PRV1"}  │ ◀─── Results ───────────── │
   │ ◀─── Tool Result ───────── │                            │
   │      {anomalies: [...]}    │                            │

MCP Implementation (packages/mcp-server/):

• Implements the @modelcontextprotocol/sdk server specification
• Supports both stdio (CLI) and HTTP streaming transports
• Routes tool calls to handler functions in packages/tools/
• Returns structured JSON responses

A2A (Agent-to-Agent Protocol)

A2A is an open protocol by Google that enables AI agents to communicate and delegate tasks. The Detection Agent (Claude) delegates deep-dive investigations to the Investigation Agent (Gemini) via A2A.

Delegation flow:

API Server                  Detection Agent              Investigation Agent
    │                          (Claude)                       (Gemini)
    │                             │                              │
    │ ── A2A Message ───────────▶ │                              │
    │    "Investigate PRV51234"   │                              │
    │                             │                              │
    │                             │ (runs initial analysis)      │
    │                             │                              │
    │ ◀── Status: "working" ───── │                              │
    │ ◀── Event: tool_call ────── │                              │
    │                             │                              │
    │                             │ ── A2A Delegation ─────────▶ │
    │                             │    "Deep investigation       │
    │                             │     needed for PRV51234"     │
    │                             │                              │
    │ ◀── Event: delegation ───── │                              │
    │                             │ ◀── Investigation Result ─── │
    │                             │                              │
    │ ◀── Final Report ────────── │                              │

A2A Implementation (agents/src/):

Each agent exposes A2A-compliant endpoints:

• /.well-known/agent.json - Agent card (capabilities, skills, authentication)
• /a2a/jsonrpc - JSON-RPC message transport
• /a2a/rest - REST streaming transport for real-time updates

Agent Cards:

Both agents publish "agent cards" describing their capabilities:

{
  "name": "Fraud Detection Agent",
  "description": "Orchestrates healthcare fraud investigations",
  "skills": [
    {
      "name": "investigate",
      "description": "Run fraud investigation on a provider"
    }
  ],
  "supportedProtocols": ["a2a/1.0"]
}

Key Components

Fraud Detection Tools

All tools are accessible via MCP and can be called by any agent:

Data

This demo uses the Healthcare Provider Fraud Detection Dataset from Kaggle, containing:

• 558,000+ healthcare claims with diagnosis codes, procedure codes, and reimbursement amounts
• 5,410 healthcare providers with fraud labels
• 138,000+ Medicare beneficiaries with demographics and chronic condition flags

The data is loaded into PostgreSQL and includes computed fields for risk scoring and anomaly detection.

Project Structure

solarys-fraud-detection-demo/
├── agents/                         # AI Agent implementations
│   └── src/
│       ├── fraud-detection/        # Claude agent (primary)
│       │   ├── server.ts           # A2A server setup
│       │   ├── claude.ts           # Anthropic SDK integration
│       │   └── skills.ts           # Agent skill definitions
│       └── claims-investigation/   # Gemini agent (specialist)
│           ├── server.ts           # A2A server setup
│           └── gemini.ts           # Google AI SDK integration
│
├── apps/
│   ├── api/                        # Express.js API server
│   │   └── src/
│   │       ├── routes/chat.ts      # Main chat endpoint (SSE)
│   │       └── lib/a2a-client.ts   # A2A SDK client wrapper
│   └── web/                        # Next.js frontend
│       └── src/
│           ├── app/trs-demo/       # Demo page
│           └── components/
│               └── investigation/  # Investigation UI components
│
├── packages/
│   ├── db/                         # Database layer
│   │   └── prisma/
│   │       ├── schema.prisma       # Data model
│   │       └── seed.ts             # Data loading
│   ├── mcp-server/                 # MCP protocol server
│   │   └── src/
│   │       ├── http.ts             # HTTP transport
│   │       └── stdio.ts            # CLI transport
│   ├── tools/                      # Tool implementations
│   │   └── src/handlers/           # Individual tool handlers
│   └── types/                      # Shared TypeScript types
│
└── data/                           # Healthcare claims CSV files

License

MIT

Contact

Engineered by the Solarys Group - AI automation and data processing.