ClaimSightv0.4.2

The multi-modal, multi-agent claims adjudication copilot.

Ingest a messy insurance claim — scanned PDF, phone photo, recorded call, policy table — and a 12-agent LangGraph workflow triages it, assesses damage from imagery, verifies coverage against the policy via RAG, scores fraud, and drafts a cited settlement recommendation. Shipped with a CI-gated eval harness and Langfuse tracing.

Agents in graph
12
supervisor + workers
Triage time
6 min
down from 45 min
Citation groundedness
94%
Ragas faithfulness
HF tasks used
18
across 4 modalities
iHow to use ClaimSight
1
Pick a claim
Click any claim card below — these are pre-loaded sample insurance claims.
2
Run the pipeline
Hit the green "Run Pipeline" button that appears — 9 AI agents start working.
3
Watch agents live
Agents stream results in real time, ending with a cited approve/deny decision.
💡 This is a workflow automation tool, not a chatbot — there's no text input. The "input" is selecting a claim and clicking Run.

Claims Queue

Live claims awaiting adjudication. Click a claim to select it →

0 claims · 3 lines of business
Loading claims...

No claim selected yet

Click a claim card above (look for the START HERE badge) — or build a custom claim below — then hit Run Pipeline to watch the 9 agents work.

Waiting for claim selection

Build Your Own Claim

Enter your own claim data — narrative, documents, photo descriptions — and run the pipeline on it.

QUICK TEMPLATES — click to auto-fill
Loading eval harness...

System Design

Production-grade patterns: not an API wrapper — orchestration, RAG, multimodal, eval, LLMOps.

Multi-Agent Pipeline Flow
1 · Ingest
Claim Bundle (PDF + photo + audio + tables)
2 · Route
Supervisor → Intake Router (zero-shot classify)
parallel fan-out
3 · Parallel Workers
Doc Extractor (NER + Doc QA)
Policy Retriever (RAG)
Vision Assessor (Det+Seg+Classif)
converge
4 · Assess
Fraud Detector (text + tabular)
Coverage Verifier (clause match)
5 · Adjudicate
Adjudicator (cited recommendation)
6 · QA Gate
Auditor (groundedness + compliance) → pass / retry / escalate
human-in-the-loop
7 · Output
Settlement memo + citations → adjuster review → payout
Agent Orchestration
  • LangGraph state machine with typed shared state channel
  • Supervisor + worker topology with parallel fan-out
  • Conditional edges (LOB + severity → path)
  • Human-in-the-loop interrupts (LangGraph checkpoint)
  • Per-agent retry + circuit breakers
RAG & Retrieval
  • Hybrid: pgvector dense (bge-m3) + BM25 sparse
  • Reciprocal rank fusion
  • Cross-encoder re-ranking (Text Ranking) — top 20 → top 5
  • Hierarchical chunking + parent-doc retrieval
  • Metadata-filtered semantic search (policy#, state, LOB)
Multimodal Vision
  • Image Classification (severity) + Object Detection (components)
  • Image Segmentation → damage-area % → cost estimate
  • Zero-Shot Image Classification (LOB routing at inference)
  • Image Feature Extraction → visual RAG / fraud dedup
  • Depth Estimation (drone roof, stretch V2)
LLMOps & Eval
  • Langfuse tracing — every span, token, cost, prompt version
  • Prompt versioning in Git, loaded by version tag
  • CI-gated eval harness (Ragas + DeepEval + LLM-judge)
  • Model routing: vLLM open models + frontier-for-reasoning (−63% $/claim)
  • NeMo Guardrails — PII redaction, schema validation
  • Semantic cache (Redis) for repeat policy queries
Recommended Tech Stack
OrchestrationLangGraph
RAG / IngestLlamaIndex + LlamaParse
Embeddingsbge-m3 (open)
Vector Storepgvector (Postgres)
LLM ServingvLLM + HF Inference
EvalRagas + DeepEval
ObservabilityLangfuse (self-hosted)
GuardrailsNeMo Guardrails
BackendFastAPI
FrontendNext.js + shadcn/ui
Queue / CacheCelery + Redis
InfraDocker + Terraform/AWS
Real-World Data APIs
  • FEMA Disaster Declarations — catastrophe flagging
  • NOAA / NWS Weather — hail/wind/flood verification
  • NHTSA Vehicle API — VIN decode + recall lookup
  • openFDA — medical/health-adjacent claims
  • USGS Flood Hazard Layer — flood-zone verification
  • Synthea — synthetic patient data
  • Kaggle fraud datasets — labeled eval corpus
LLMOps — Production Maturity
Observability

Langfuse: every span, token cost, latency, prompt version traced

Prompt Versioning

Prompts in Git, A/B via feature flag, rollback in seconds

Model Routing

Open models for extraction, frontier for reasoning — −63% $/claim

Guardrails

NeMo: PII redaction (SSN, VIN), output schema validation