ClaimGuard: Multi-Agent Insurance Claims System
ClaimGuard is a multi-agent auto-insurance claims system that treats data governance as a first-class collaboration problem, not an afterthought. Four agents — Intake, Assessment, Payout, and a Compliance Firewall — communicate entirely through Band. Most multi-agent demos pass full records down a fixed pipeline, leaking sensitive data (SSNs, medical details, bank accounts) to agents that never needed it. ClaimGuard inverts this: the Compliance Firewall is a queryable Band participant, not a preprocessing step. Before Assessment or Payout act on any claim data, they actively request clearance over Band — "I am Assessment, clear me for my role" — and the Firewall responds with only the fields that agent is permitted to see, plus a logged justification for every decision. The firewall itself is a deterministic policy engine, not an LLM—compliance decisions must be auditable and immune to prompt injection, not probabilistic. Intake extracts structured claims from free text and validates policy status; Assessment runs a RAG-powered Four-Point Coverage Analysis against real policy documents to approve, deny, or escalate; Payout simulates disbursement. Every firewall decision is logged with a field-level reason, producing a complete data-minimization audit trail
Category tags:
.png&w=96&q=75){kind=small}
.png&w=96&q=75){kind=small}
"Application of technology: Four insurance claims agents (Intake/free-text extraction + policy validation, Assessment/RAG Four-Point Coverage Analysis against real policy docs, Payout, Compliance Firewall/deterministic policy engine) with active clearance request pattern — agents @mention Firewall requesting field-level clearance before accessing data. Firewall responds with permitted fields + logged justification. Chroma RAG + HuggingFace. Band Agentic Mesh. Presentation: Outstanding data governance innovation — "Compliance Firewall as a queryable Band participant, not a preprocessing step." Active clearance pattern (agents request permission before access) inverts the typical pipeline-based data leakage pattern. Field-level justification logging creates a complete data-minimization audit trail. Three-person team. Business value: Insurance claims processing with active data-minimization compliance (agents only see what they need, with logged justification for every access decision) directly addresses GDPR/CCPA data minimization requirements for regulated financial workflows. Originality: The active clearance request pattern — agents requesting field-level permission from a deterministic Compliance Firewall via Band before accessing claim data — is a genuinely original data governance architecture for multi-agent systems. This inverts the typical pipeline model and makes Band the access control layer, not just the coordination layer."
Sanem Avcil
For Innovators & Creators