Current project

FeaturedResearch Prototype

DocPipeline

View project

AI-powered document extraction system that turns PDFs and images into structured outputs through OCR, schema-driven extraction, and validation.

Problem

Raw OCR output is not enough for business workflows. Teams still need to interpret fields, normalize formats, and move extracted data into systems they already use.

Architecture

DocPipeline is built as a modular, multi-stage system: document input, pluggable OCR, extraction logic, schema-based shaping, and async export into downstream integrations.

Impact

The project demonstrates a cleaner path from documents to usable structured data across receipts, purchase orders, packing slips, utility bills, expense reports, and quotes.

StackPaddleOCRAzure Document IntelligenceLLM ExtractionSchema ValidationAsync WorkersWebhooks
Related readingLearning Log
Designing Reliable AI Pipelines

Reliability in AI systems is mostly about interfaces, observability, fallback behavior, and operational discipline — not the model.

Learning Log
I Spent Months Improving OCR Accuracy. The Real Breaks Were Somewhere Else.

Schema validation, queue design, and human-in-the-loop are the three layers where production document AI systems actually break — not the model.

Learning Log