Pioneers AI Hackathon 2025

FidelityAI

Autonomous Insurance Fraud Detection

$100B annual fraud losses. 2-3 weeks manual review per claim.

We built a solution.

Voice Interview - AI Fraud Detection
Voice Interview Technology
Powered by ElevenLabs AI
10
AI Agents
10
Minutes
0
Human Input
0.80
Fraud Risk
Press Space or ↓ to continue

10-Agent Autonomous Pipeline

End-to-end fraud detection with zero human intervention

Phase 1: Data Ingestion
1
Document
Gemini Flash
2
Image Analysis
Gemini Vision
3
Voice/Text
Gemini Flash
↓
Phase 2: Knowledge Retrieval
4
Guidelines Search
Gemini Pro + Qdrant
5
Similar Claims
Gemini Pro + Qdrant
↓
Phase 3: Voice Interview
6
Question Generator
Gemini Pro
7
Voice Interviewer
ElevenLabs AI
↓
Phase 4: Adversarial Analysis
8
Agreement
Pro Approval
9
Disagreement
Anti Approval
↓
Phase 5: Final Verdict
10
Final Judge
APPROVE • DENY • INVESTIGATE
Stateless Agents • File-Based Communication • Full Audit Trail

Technology Stack

3 technologies powering autonomous fraud detection

Google Gemini

Gemini 2.5 Flash
Fast ingestion • Agents 1-3
Gemini 2.5 Pro
Deep reasoning • Agents 4-10
text-embedding-004
768D vectors
Powers
10 Agents

Qdrant

Insurance Guidelines
60 policies • COSINE
Historical Claims
180 claims • Hybrid search
Pattern Matching
Dense + Sparse vectors
Search Speed
<2 seconds

ElevenLabs

Conversational AI
Real phone calls
Behavioral Analysis
Detect hesitation • evasion
Outbound Calling
24/7 availability
Agent 7
Voice Interviewer
Integration Flow
Gemini → Qdrant → ElevenLabs → Gemini

Live Case: CLM-2024-08472

Real fraud detection from our autonomous system

Claimant
Sarah Mitchell
Incident
Rear-end collision
Vehicle
2022 Toyota Camry
Location
Austin, TX
📄

Step 1: Claim Submitted

Agent 1
Rear-ended at red light. Police report #APD-2024-11083. Witnesses provided. Vehicle: 2022 Toyota Camry, License: XYZ-7890
✓ Initial claim appears legitimate
🚨

Step 2: Red Flag Detected!

Agent 2
Image analysis reveals vehicle mismatch:
Claimed Vehicle
2022 Toyota Camry
License: XYZ-7890
Photos Show
2022 Lexus ES 350
Different vehicle!
🚨 MATERIAL MISREPRESENTATION
🔍

Step 3: Qdrant Pattern Matching

Agent 5
Found 3 similar fraud patterns in 180 historical claims:
● CLM-00014 - Inconsistent damage (fraud_detected)
● CLM-00024 - Wrong vehicle photos (fraud_detected)
● CLM-00026 - Houston fraud ring (fraud_detected)
🎤

Step 4: Voice Interview

Agent 7 (ElevenLabs)
Agent: "Why did you submit photos of a Lexus?"
Claimant: "I think I forgot... I don't remember that much"
Agent: "Is all the damage in these photos from this accident?"
Claimant: "Yes" [LIE DETECTED]
Agent: "Let me verify the police report..."
Claimant: [CALL TERMINATED]
🚨 Evasive behavior + Failure to cooperate
⚖️

Step 5: Final Verdict

Agent 10
Agreement Agent Score
0.35
"Strong initial docs, possible mistake"
Disagreement Agent Score
0.95
"Clear fraud, deceptive behavior"
FINAL DECISION
DENY
Fraud Risk: 80%
Material misrepresentation + Failure to cooperate
Complete Evidence Trail
10 Markdown Files • Human-Readable • Fully Auditable

Impact & Results

This is an autonomous AI insurance company

⚡
95% Faster
45 minutes vs 2-3 weeks manual review
💰
$0.10
Per claim vs $500+ manual investigation
🎯
100%
Explainable decisions with full audit trail
🤖
0
Human intervention until final review

Autonomous AI Company

✓ End-to-End Automation
From claim submission to final verdict
✓ Multi-Agent Orchestration
10 specialized agents working together
✓ Adversarial Analysis
Reduces bias through structured debate
✓ Voice Engagement
Real conversations, not text chatbots
✓ Pattern Recognition
Learns from 180 historical cases
✓ Explainable AI
Every decision fully documented
The Future is Autonomous
One-person AI companies that scale to millions
View GitHub Repo
Built with: Google Gemini • Qdrant • ElevenLabs • Google ADK • Next.js • FastAPI