Case study

AI Insurance Claims System

RAG, OCR, and workflow automation for regulated claims — case study

End-to-end narrative of how we moved from manual document triage to grounded AI assistance with auditability. Names and figures are representative; tune copy to what you can disclose.

Timeline

Weeks 1–3
Discovery
Mapped analyst journeys, policy corpus sources, and compliance checkpoints. Defined non-negotiables: citations, retention, PII boundaries.
Weeks 4–10
MVP pipeline
OCR + normalization workers, vector index over policy chunks, rules engine for hard gates, LLM path only when retrieval confidence cleared a bar.
Weeks 11–16
Hardening
Evaluation sets for citation accuracy, cost caps per tenant, DLQs and replay for workers, dashboards for queue depth and model spend.
Weeks 17+
Rollout
Phased enablement by LOB, human-in-the-loop for low-confidence, feedback loop into chunking and re-ranking.

Request / response flow

Analyst UI          API              Workers           Vector DB / LLM
   |                 |                  |                    |
   |-- upload doc -->|                  |                    |
   |                 |-- enqueue OCR -->|                    |
   |                 |                  |-- embed chunks --->|
   |                 |<-- status -------|                    |
   |<-- draft -------|                  |                    |
   |                 |-- RAG query ------------------------>|
   |                 |<-- answer + citations ----------------|

Deeper architecture

  ┌────────────┐    ┌─────────────┐    ┌──────────────────┐
  │   Gateway  │───▶│ Claims API  │───▶│  Domain services │
  │  + authZ   │    │  (ASP.NET)  │    │  + outbox        │
  └────────────┘    └──────┬──────┘    └────────┬─────────┘
                           │                    │
                    ┌──────▼──────┐      ┌──────▼─────────┐
                    │  Bus /    │      │ PostgreSQL      │
                    │  queues   │      │ + vectors       │
                    └──────┬──────┘      └────────────────┘
                           │
                    ┌──────▼──────┐      ┌────────────────┐
                    │ OCR / NLP  │      │ LLM router     │
                    │ workers    │      │ (guardrailed)  │
                    └────────────┘      └────────────────┘

Modal snapshot (same as home)

Problem, decisions, and metrics also live on the main portfolio card — this page extends with timeline and evolution.

Claims analysts were drowning in unstructured documents and policy lookups; decisions were slow, inconsistent, and hard to audit for compliance.

What I'd change next

Tighter offline eval loops: golden sets per product line with regression gates in CI.
Multi-model routing: cheaper models for classification, premium only for synthesis.
Stronger lineage: document version → chunk → embedding id in every audit record.
Federated search if policies span multiple repositories with different ACLs.

Demo

Demo · Simulated RAG response

Press run to simulate retrieval → answer.