Bihar Government Kaithi Script OCR & Digitalization Platform
End-to-end OCR system and web platform for digitizing degraded historical Kaithi script documents - built for Bihar Government and Bhashini.
Project Snapshot
Role
Project Lead - full end-to-end ownership, leading a team of 10 people
Organization
Bihar Government + Bhashini (Ministry of Electronics & IT)
Timeline
2025 - Active Development
Status
Active Development
Deployment
Planned on Bhashini / Government infrastructure
Tech Stack
Impact & Results
- Digitizing irreplaceable historical documents from Bihar
- Building the first OCR system for Kaithi script
- Full-stack web platform for search, browse, and export of digitized text
An end-to-end OCR system and web platform for digitizing degraded historical Kaithi script documents, built for the Bihar Government and Bhashini. I led the complete OCR pipeline design and full-stack web platform development.
The Problem
Thousands of historical documents written in Kaithi script (an ancient Indo-Aryan script historically used in Bihar and UP for administrative records) are deteriorating. These documents contain irreplaceable historical, legal, and cultural records. The government needed to digitize these documents, make them searchable, and build an OCR system capable of reading degraded manuscripts.
The Solution
A complete digitalization workflow involving data collection, custom annotation tooling, a multi-stage OCR pipeline (detection + recognition), and a full-stack web platform for browsing, searching, and exporting digitized text.
Architecture & Implementation
OCR Pipeline:
*Preprocessing:* Denoising, binarization, deskewing.
*Detection:* CRAFT / DBNet to find text regions.
*Recognition:* CRNN / TrOCR fine-tuned specifically on Kaithi script.
*Post-processing:* Language model correction.
Web Platform: End-to-end digitalization web platform built with Next.js and Django for search, browse, and export of digitized text.
Management: Designed the annotation workflow and led a team of 10 people for data collection and model evaluation (metrics: CER, WER, mAP).
Data Used: Custom dataset of historical government documents from Bihar (highly degraded, significant handwriting variation).
Tech Stack: Python, PyTorch, TensorFlow, OpenCV, CRAFT, DBNet, CRNN, TrOCR, HuggingFace, Next.js, Django, PostgreSQL.
Challenges & Solutions
Data Scarcity: Complete lack of pre-existing annotated datasets for the historical Kaithi script.
Data Quality: Training robust vision models on highly degraded, noisy historical manuscripts.
Workflow: Managing a 10-person annotation team and ensuring data quality.
Key Takeaways
End-to-end management of a complex AI lifecycle—from data collection and custom annotation tool building to model deployment.
Advanced computer vision techniques for historical document processing.
Team leadership and project management.