Hebrew OCR Pipeline
Zero-hallucination OCR pipeline processing thousands of multi-column historical Hebrew texts at 96% accuracy.
Project Overview
A zero-hallucination Hebrew OCR pipeline built at Introva, processing thousands of multi-column historical texts at 96% accuracy. Combines Google Vision API with custom layout-detection algorithms, an advanced post-processing engine backed by a 23,000+ word lexical database, and Sefaria API integration for citation normalisation.
Key Features
- 96% accuracy on multi-column historical Hebrew texts
- Custom layout-detection algorithms for columns, footnotes, and marginal notes
- 23,000+ word lexical database for word-level correction and disambiguation
- Zero-hallucination design — surfaces low-confidence tokens rather than inventing text
- Sefaria API integration for citation validation and normalisation
Case Study
Context
Built at Introva for a corpus of thousands of multi-column historical Hebrew texts with pre-modern orthography, mixed scripts, and variable scan quality.Pipeline
1. Image preprocessing — deskew, binarise, denoise, consistent DPI 2. Layout detection — custom logic for columns, footnotes, and marginal notes before OCR 3. Raw OCR — Google Vision API for character recognition 4. Lexical post-processing — 23,000+ word database for correction, context-aware disambiguation, abbreviation expansion, and Sefaria API citation normalisationResults
Related Projects
Smart Hire AI
AI-powered recruitment platform with active enterprise pilots. Best FYP Award — UMT Software Engineering Department 2026.
Vibe Terminal
Cross-platform AI terminal assistant — execute shell commands from natural language, with voice I/O and multi-step agent workflows.
Smart Grade AI
AI-powered LMS with RAG-based mentoring — Microsoft Imagine Cup 2025 Global Semifinalist, backed by $25K Azure credits from Microsoft for Startups, in institutional rollout at UMT.