Flagship Product #2
Agentic OCR
Document Intelligence System. VLM-powered extraction that replaced a broken regex pipeline with 100% accuracy on production, at 35x lower cost.
100%
Production Accuracy
$0.01
Per Document
16s
Avg Processing
0%
Failure Rate
Product Walkthrough
See Agentic OCR in Action
Full demo of the document intelligence pipeline processing real purchase orders and acknowledgements on production.
Production Benchmarks
Before vs After
Tested on 15 documents across 4 different approaches over 2 weeks. Production verified on 5 live documents.
Header Accuracy
40%
100%
127/127 fields
Cost Per Document
$0.03-0.05
$0.007
35x cheaper
Processing Speed
52 seconds
8-12s
5x faster
Failure Rate
73%
0%
Zero failures
Docs Processed
4/15
15/15
100% coverage
Pipeline Services
3 services
1 service
Simplified stack
R&D Journey
3 Approaches Compared
We tested 4 approaches over 2 weeks. Here's how they stacked up.
Attempt 1
Regex Parser
Headers40%
Line Items90%*
Cost/Doc$0.03-0.05
Speed52s
Failure Rate73%
Docs Tested4/15
Attempt 2
LLM Mapper (Haiku 3)
Headers98%
Line Items80%
Cost/Doc$0.03-0.05
Speed97s+
Failure Rate73%**
Docs Tested9/15
Attempt 3
VLM Direct (Haiku 4.5)
Headers100%
Line Items73.6%
Cost/Doc$0.007
Speed8-12s
Failure Rate0%
Docs Tested15/15
How It Works
The Pipeline
PDF in, structured JSON out. No OCR, no regex, no intermediary. One service call.
PDF Input
Any format
PDF to PNG
pymupdf, 150 DPI
Claude Haiku 4.5
Bedrock tool use
Validation Engine
5 rule groups
Structured JSON
Schema-validated
Under the Hood
Architecture Overview
What we removed, what we built, and how it maps to the 5-layer architecture vision.
What Got Removed
DeepSeek OCR server (Docker/ECS)
Regex field_aliaser_service.py
OCR MCP server HTTP calls
Markdown intermediary step
What Got Added
VLMExtractionService (~300 lines)
Per-document-type schemas
Per-document-type system prompts
Validation engine (5 rule groups)
Infrastructure
ECS Fargate (us-east-1)
AWS Bedrock (Haiku 4.5)
Cross-region inference profile
Zero new dependencies
Validation Engine
Financial validation (qty x price)
Date sanity checks
Required field enforcement
Cross-field logic checks
Auto-correction (formats, nulls)
Full Deck
Product Presentation
Complete technical breakdown with benchmarks and architecture diagrams.
Deep Dive
Technical Documentation
Full VLM benchmark report and production deployment verification.
Case Study & Benchmarks
VLM Benchmark Full Report
15 documents, 4 approaches, 2 weeks of testing. Complete accuracy and cost analysis.
Production Benchmarks
5-Doc Production Verification
Live endpoint tested against 5 real vendor documents. 100% accuracy, all AUTO_APPROVED.