ClaimFlow — AI Autopilot Agent for Claims Processing

• 

  dashboard at localhost:8080

What It Does

ClaimFlow automates the entire claims processing workflow. A customer submits a claim via WhatsApp, web form, or API with text and photos as evidence. The agent processes it through six stages:

1. Intake — Normalizes messages from any channel, detects language
2. Classify — Qwen 3.7 Max identifies claim type, object, urgency, sentiment
3. Verify — Qwen Vision Max inspects image evidence against the claim
4. Fraud Detection — Four Rs framework (Recognize → Reject → Reveal → Route)
5. Decide — Auto-approve, request info, or escalate to human
6. Respond — Auto-reply in user's language or create human case

How I Built It

Built with Python 3.10+, Flask, and the Anthropic-compatible Qwen Cloud API. Two Qwen models power the pipeline: qwen3.7-max for text classification and qwen-vl-max for image evidence verification. The human-in-the-loop dashboard runs on Flask with server-side templates.

Challenges

Getting the 6-stage pipeline to handle edge cases gracefully. Building the Four Rs fraud framework to combine vision-based evidence checking with user history patterns. Making the agent fail-safe — when uncertain, it escalates rather than making a wrong decision.

What I Learned

AI agents work best when they know their limits. The autopilot handles 80% of claims, but the human dashboard for the remaining 20% is what makes it production-ready. Multi-language support is a small detail that makes a big difference for real-world deployment.

Built With

• anthropic-sdk
• flask
• pytest
• python
• qwen-cloud-(qwen3.7-max
• qwen-vl-max)