PaperExtractor: AI-Powered PDF Data Extraction

Overview

PaperExtractor is an AI-powered tool that extracts specific information from PDF documents. Upload a PDF, specify what information you need, and the application automatically finds it with source citations.

View on GitHub

Demo

I also shared this project and my thoughts on building it in a LinkedIn post

Learning Goals

• Visual Derendering: Leveraging Gemini 3 Pro’s advanced visual derendering capabilities to reverse-engineer the visual layout of a PDF back into structured code.
• Two-Stage Processing: Converting PDFs to Markdown first (preserving structural hierarchy), then extracting requested fields with proof snippets.
• Overcoming PDF Limitations: Traditional PDF extraction is notoriously unreliable since PDFs are designed for humans, not computers, resulting in misordered text, broken tables, and poor formatting. The derendering approach addresses these challenges.
• Visual Interface: Building a three-panel display showing original PDF pages, OCR text, and extracted results.

Tech Stack

• Frontend: React + Tailwind CSS
• Backend: Node.js + Express
• AI API: Google Gemini 3 Pro for intelligent extraction
• PDF Processing: pdf-to-img for document handling

Takeaways

• Two-stage processing (PDF to Markdown, then extraction) improves accuracy and allows for better source citations.
• JSON schema mode in Gemini makes structured data extraction much more reliable.
• Providing field context instructions significantly improves extraction quality for domain-specific documents.
• Building a visual interface that shows the extraction proof alongside results helps users trust and verify the output.
• Markdown is widely recommended by model providers for use with Large Language Models, making it an ideal intermediate format.