Meridian Claims — System Architecture

Meridian Claims — System Architecture An architecture diagram generated by Archify. Claims Adjuster · browser · Architecture component Claims Adjuster browser Meridian Portal · Next.js + shadcn/ui · Architecture component · queue · FNOL · audit trail Meridian Portal Next.js + shadcn/ui queue · FNOL · audit trail FastAPI Backend · uv · :8000 · Architecture component FastAPI Backend uv · :8000 SQLite · claims · findings · activity · Architecture component SQLite claims · findings · activity Policy Verification · check_coverage · Claims pipeline Policy Verification check_coverage Medical History · lookup_medical_history · Claims pipeline Medical History lookup_medical_history Fraud / Exception · fraud_score · Claims pipeline Fraud / Exception fraud_score Decision Gate · deterministic · Claims pipeline Decision Gate deterministic MCP Server · stdio tools · Tool + model boundary MCP Server stdio tools Ollama · qwen2.5:7b, local · Tool + model boundary Ollama qwen2.5:7b, local OpenRouter · free-tier fallback · Tool + model boundary OpenRouter free-tier fallback Langfuse Cloud · traces · sessions · scores · Architecture component · OpenInference/OTel Langfuse Cloud traces · sessions · scores OpenInference/OTel HTTPS REST + SSE persists claims run_claim(), sequential findings decision check_coverage lookup_history fraud_score LLM calls, primary fallback trace + score per claim Claims pipeline Tool + model boundary Legend Frontend Backend Database Cloud External

Portal

  • • Next.js + shadcn/ui claims-ops portal
  • • Live SSE activity via floating widget + audit trail

Agent pipeline

  • • CrewAI sequential process: Policy → Medical History → Fraud
  • • Deterministic decision gate, not LLM phrasing

Tools + model

  • • MCP server (stdio) exposes 3 mock claim-systems tools
  • • Ollama local primary, OpenRouter free-tier fallback

Observability

  • • Every LLM + tool call traced via OpenInference/OTel
  • • One Langfuse Session per claim, groundedness Score per finding