AIHub

Overview

AIHub is a document intelligence platform that brings AI-powered processing capabilities to the browser. It enables teams to extract text from images and PDFs, manipulate documents, and process content at scale - all without uploading sensitive documents to external servers.

The Challenge

Document processing in enterprise environments faces several constraints:

• Sensitive documents can't be uploaded to third-party services
• Manual data extraction is slow and error-prone
• Existing OCR solutions require complex server infrastructure
• Teams need to process documents in various formats
• Integration with existing authentication systems is required

Solution

A browser-based platform that:

• Runs OCR entirely in the browser using Tesseract.js
• Manipulates PDFs client-side with pdf-lib
• Processes images with Sharp for optimal OCR input
• Stores data locally with IndexedDB for privacy
• Authenticates via Azure AD for enterprise deployment

Technical Architecture

Client-Side OCR

Tesseract.js enables accurate text recognition without server round-trips:

// Initialize Tesseract worker
const worker = await createWorker('eng')

// Process image with confidence scoring
const { data } = await worker.recognize(imageData)

// Extract text with position data
const blocks = data.blocks.map(block => ({
  text: block.text,
  confidence: block.confidence,
  bbox: block.bbox
}))

PDF Processing Pipeline

Multi-stage document handling:

1. Ingestion - Parse PDF structure with pdfjs-dist
2. Extraction - Render pages to canvas for OCR
3. Enhancement - Pre-process images with Sharp
4. Recognition - Run Tesseract on enhanced images
5. Assembly - Combine results with pdf-lib

Image Pre-processing

Sharp optimizes images before OCR:

• Convert to grayscale for better recognition
• Apply adaptive thresholding
• Deskew rotated scans
• Remove noise and artifacts

Local Storage Strategy

IndexedDB provides persistent, private storage:

• Processed documents cached locally
• User preferences and settings
• Offline capability for previously processed files

Key Features

Smart Text Extraction

AI-powered OCR that handles:

• Scanned documents and photos
• Multi-column layouts
• Tables and structured data
• Handwritten text (with reduced accuracy)

PDF Manipulation

Client-side operations:

• Merge multiple PDFs
• Split documents by page
• Extract specific pages
• Add watermarks and annotations

Batch Processing

Queue multiple documents for processing:

• Progress tracking per document
• Parallel processing with Web Workers
• Resume interrupted batches

Enterprise Authentication

Azure MSAL integration:

• Single sign-on with company credentials
• Token-based session management
• Automatic token refresh

Technical Stack

• Nuxt 3 for the application framework
• Tesseract.js 6 for browser-based OCR
• pdf-lib for PDF creation and manipulation
• pdfjs-dist for PDF rendering and parsing
• Sharp for image processing
• Canvas API for image manipulation
• IndexedDB (idb) for local storage
• Azure MSAL for authentication
• Playwright for E2E testing

Performance Optimizations

Document processing is resource-intensive. Key optimizations:

• Web Workers - OCR runs off the main thread
• Progressive loading - Process visible pages first
• Caching - Store intermediate results
• Lazy initialization - Load Tesseract only when needed
• Memory management - Release resources after processing

Privacy by Design

All processing happens in the browser:

• Documents never leave the user's device
• No server-side storage of content
• Authentication tokens stored securely
• Clear data option for sensitive sessions

Results

• Processes 50+ page documents in under a minute
• 95%+ accuracy on clean printed text
• Zero server infrastructure for document processing
• Deployed to enterprise teams with strict data policies

Lessons Learned

1. Browser capabilities are impressive - Modern browsers can handle serious workloads
2. OCR quality depends on input - Image pre-processing is crucial
3. Memory limits are real - Large documents need careful chunking
4. UX during processing - Users need feedback for long operations
5. Offline-first wins - Local storage makes the app feel instant