SafeClaim AI: Multi-Agent Insurance Claims Adjuster on AWS EKS

The Challenge

Property and casualty insurance claims required 3 to 5 business days of manual adjuster review per claim. A 23% fraud rate was slipping through because human reviewers lacked the tooling to cross-reference thousands of historical cases simultaneously. The existing system had no observability, no horizontal scaling, and zero auditability on claim decisions.

Our Solution

Sciensify designed and built a three-agent LangGraph pipeline deployed on Amazon EKS. The Ingestor Agent uses a multimodal Vision LLM (Claude Sonnet and GPT-4o) to extract structured damage metadata from accident photos with Pydantic validation. The Fraud Detective Agent embeds each claim and runs cosine similarity search against 15,420 historical fraud cases stored in a Pinecone vector database, flagging duplicate phone numbers, repeat claimants, and suspiciously similar incident patterns. The Adjuster Agent retrieves relevant policy clauses via RAG over the customer PDF and generates a payout recommendation with confidence scoring. Claims above 0.80 confidence auto-approve and write to Amazon RDS. Claims below threshold pause and route to a Human-in-the-Loop Next.js dashboard for adjuster review. The entire infrastructure was provisioned with Terraform across an AWS VPC with public and private subnets, EKS with HPA autoscaling, S3 for photo storage, ECR for container images, and GitHub Actions plus ArgoCD for GitOps continuous delivery. Prometheus and Grafana provide full observability across latency, queue depth, and cost per claim.

The Results