All Case Studies
InsuranceGlobal Insurer2024
Lumen Document Intelligence
Replaced a 90-person document review team's intake bottleneck with a multi-model OCR + LLM pipeline that handles 14 document types with full audit trails.
Client
Global Insurer
Sector
Insurance
Duration
18 weeks
Team
3 engineers · 1 product manager
The Challenge
The starting point.
Claims documents arrived in 11 languages and 14 formats. The existing OCR vendor produced 76% accuracy and the review team was 5 weeks behind on backlog.
The Approach
What we built.
- 01Benchmarked 4 vision-language models on a stratified sample
- 02Designed a confidence-routing pipeline with human-in-the-loop fallback
- 03Built a regulator-friendly audit log capturing every model decision
- 04Migrated reviewers from data-entry to exception-handling roles
Results
What shipped, what changed.
94%
End-to-end extraction accuracy
8x
Throughput vs. legacy OCR
0
Backlog days, post-launch
11
Languages supported
Stack
What's running in production.
ClaudeGPT-4oAWS TextractPostgresTemporalDatadog
Next case study
Helix RAG Search