Architecture of IntelligenceDrawing set · Akshay Bajpai

W-410Works

Insurance Document Intelligence on AWS

Multimodal RAG, Claude and GPT-4 Vision, LayoutLMv3 extraction, and ensemble underwriting models: 95%+ accuracy at 10K+ queries/day.

Scale
1:1
Rev
A
Issued
Jun 1, 2024
Reading
2 min
Engagement
Insurance · document AI
  • AWS Bedrock
  • RAG
  • Claude
  • GPT-4 Vision
  • LayoutLMv3
  • LightGBM
  • Textract
  • Python
Schedule of outcomes3 items
  1. 0195%+extraction accuracy
  2. 0278%underwriting efficiency gain
  3. 0310K+queries/day

Problem

Insurance operations ingest complex semi-structured documents: policies, endorsements, scans with tables and handwriting. Manual extraction does not scale; naive OCR misses layout; generic chatbots invent coverage details. The platform needed multimodal extraction, retrieval-grounded Q&A, and underwriting models that improve decisions without bypassing human sign-off.

Architecture

Document intelligence pipeline:

  • RAG with Claude 3.5 Sonnet and GPT-4 Vision for multimodal understanding
  • LayoutLMv3 for layout-aware extraction on challenging pages
  • OCR-to-RAG path using AWS Textract and Apache Tika for high-volume ingestion

Deployed on AWS Bedrock with hybrid retrieval, reranking, prompt optimization, and automated evaluation loops.

Downstream ML: LightGBM, CatBoost, and XGBoost ensembles that improved underwriting efficiency by ~78% on the workflows we targeted.

Scale & accuracy

  • 95%+ extraction accuracy across high-volume ingestion pipelines
  • 10K+ queries/day sustained on the RAG serving path

Lessons

  1. Layout is signal: vision + layout models beat text-only pipelines on insurance PDFs.
  2. Evaluation pipelines are production code: prompt tweaks without regression tests erode the 95% claim.
  3. Ensembles still win tabular underwriting: LLMs extract; gradient boosting decides when features are structured.

Related reading

Architecture Design for Healthcare AI: shared themes on auditability and human override, applied across regulated domains.

W-410record

sheet
W-410
title
Insurance Document Intelligence on AWS
subtitle
Multimodal RAG, Claude and GPT-4 Vision, LayoutLMv3 extraction, and ensemble underwriting models: 95%+ accuracy at 10K+ queries/day.
discipline
W · Works
scale
1:1
revision
A
issued
Jun 1, 2024
refs
none
series
Works
words
201
stack
AWS Bedrock, RAG, Claude, GPT-4 Vision, LayoutLMv3, LightGBM, Textract, Python
metrics
95%+ extraction accuracy · 78% underwriting efficiency gain · 10K+ queries/day

sourcemarkdown


## Problem

Insurance operations ingest **complex semi-structured documents**: policies, endorsements, scans with tables and handwriting. Manual extraction does not scale; naive OCR misses layout; generic chatbots invent coverage details. The platform needed **multimodal extraction**, **retrieval-grounded Q&A**, and **underwriting models** that improve decisions without bypassing human sign-off.

## Architecture

**Document intelligence pipeline:**

- RAG with Claude 3.5 Sonnet and GPT-4 Vision for multimodal understanding
- LayoutLMv3 for layout-aware extraction on challenging pages
- OCR-to-RAG path using **AWS Textract** and Apache Tika for high-volume ingestion

**Deployed on AWS Bedrock** with hybrid retrieval, reranking, prompt optimization, and automated evaluation loops.

**Downstream ML:** LightGBM, CatBoost, and XGBoost ensembles that improved **underwriting efficiency by ~78%** on the workflows we targeted.

## Scale & accuracy

- **95%+ extraction accuracy** across high-volume ingestion pipelines
- **10K+ queries/day** sustained on the RAG serving path

## Lessons

1. **Layout is signal**: vision + layout models beat text-only pipelines on insurance PDFs.
2. **Evaluation pipelines are production code**: prompt tweaks without regression tests erode the 95% claim.
3. **Ensembles still win tabular underwriting**: LLMs extract; gradient boosting decides when features are structured.

## Related reading

[Architecture Design for Healthcare AI](/blog/healthcare-ai-architecture/): shared themes on auditability and human override, applied across regulated domains.