FidelityAI
Team consisting of Safouane Chergui (Doctolib Lead Data Scientist; RAG/LLM, Python/R), Achraf Ez (CTO, LumiereAI; generative video), Imad Ezzejjari (AI founder; GCP, Langchain).
YouTube Video
Project Description
FidelityAI - Autonomous Insurance Fraud Detection
Core Problem
Insurance companies lose billions annually to fraud. Manual claim reviews are slow, inconsistent, and miss sophisticated patterns. Our solution automates the processing of the claims and detects frauds.
Solution
FidelityAI : A 10-agent autonomous system that detects fraud through voice interviews, adversarial analysis, and historical pattern matching—delivering decisions in 10 minutes vs. days.
Target Customer
Insurance companies processing high-volume claims.
Key Features
• Voice Interviews: ElevenLabs Conversational AI conducts natural claim interviews
• Adversarial Analysis: 6 specialized agents debate pro/con evidence before final judgment
• Pattern Detection: Qdrant vector database with hybrid search (semantic + keyword) flags similar historical fraud cases across many claims in <2 seconds
• 10-Minute Pipeline: Autonomous end-to-end processing with zero human intervention
• Database-Driven: All evidence, analysis, and decisions stored in Qdrant for complete audit trail and continuous learning
Demo Flow
Upload claim → AI interviews claimant → Multi-agent analysis → Fraud verdict with evidence
Technologies
Google Gemini 2.5 (Flash/Pro), ElevenLabs Conversational AI, Qdrant Vector DB, Google ADK, FastAPI SSE
Qdrant Vector Database Integration
• 2 Collections: Insurance policies (60 docs) + Historical claims (180+ docs)
• Hybrid Search: 70% semantic similarity + 30% keyword matching
• Real-Time: <2 second search across entire claims database
• Pattern Detection: Detects fraud rings through geographic clustering and temporal patterns
• Continuous Learning: Each processed claim automatically indexed for future pattern matching
• Audit Trail: Complete storage of all agent outputs, evidence, and decisions
Agent Logic
Sequential pipeline: Inputs (Documents and photos) Analysis → Policy Retrieval (Qdrant) → Interview Prep → Voice Conductor → Pro/Con Agents (6) → Final Judge → Report Generation. Each agent is stateless, storing all results in Qdrant database for transparency & audit compliance.
Judging Criteria Alignment
• Innovation: First autonomous voice-based fraud detection pipeline
• Impact: Reduces processing time by 99% (10 min vs. 3 days)
• Technical: Multi-agent orchestration with adversarial reasoning
• Feasibility: Frontend is ready, Tested with real claim data
Prior Work
Nothing