AutoClaims — Multi-Agent AI Insurance Processing

5 AI agents powered by Qwen Cloud automate claims end-to-end: photo intake, validation, assessment, review gate with human oversight, and resolution with real-time pipeline UI.

Comment

Inspiration

Insurance claims processing is traditionally slow, paperwork-heavy, and prone to human error. Small claims can take weeks, while fraudulent claims slip through. We built AutoClaims to demonstrate how 5 specialized AI agents working together can process an insurance claim end-to-end in seconds — from photo upload to payout — while keeping a human in the loop for high-risk cases.

What it does

AutoClaims automates the entire claims lifecycle through 5 AI agents:

  1. Intake Agent (Qwen3.7-Plus) — Extracts structured claim data from submissions and analyzes damage photos using vision AI
  2. Validation Agent (Qwen3.7-Max) — Verifies policy validity, coverage, and claimant identity
  3. Assessment Agent (Qwen3.7-Max) — Estimates repair costs, evaluates damage severity, and assesses fraud risk
  4. Review Gate (Qwen3.7-Max) — Decides if human review is needed based on payout amount and risk score
  5. Resolution Agent (Qwen3.6-Flash) — Generates the final approval/rejection letter with payout details

The system features a real-time React Flow pipeline visualization, a drag-and-drop photo upload with AI vision analysis, and a human review dashboard for operator oversight.

How we built it

  • Backend: Python FastAPI with SQLite, organized into 5 agent modules orchestrated by a supervisor
  • Frontend: Next.js 15 with TypeScript, Tailwind CSS, and React Flow for the pipeline UI
  • AI: Qwen Cloud API (qwen3.7-max for reasoning, qwen3.7-plus for vision, qwen3.6-flash for speed)
  • Deployment: Docker multi-stage builds, ready for Alibaba Cloud ECS
  • Photo Analysis: Qwen3.7-Plus vision API classifies damage type, severity, and estimates repair costs

Challenges we faced

  • Qwen API integration: Getting chat_json to work reliably with structured outputs required careful prompt engineering and error handling
  • Pipeline state management: Coordinating 5 agents with shared state and persistent progress tracking
  • Windows compatibility: Turbopack native bindings broken on Windows — used --webpack fallback for Next.js builds
  • Frontend-backend synchronization: Ensuring the real-time pipeline visualization accurately reflects processing state
  • Cross-origin issues: CORS configuration between the frontend (port 3010) and backend (port 8000)

What we learned

  • How to chain multiple AI agents with different model specializations
  • Qwen Cloud's vision API capabilities for real-world image analysis
  • Building human-in-the-loop workflows with AI decision gates
  • Creating intuitive pipeline visualizations with React Flow
  • Containerization strategies for multi-service AI applications

Built With

Share this project:

Updates