Integrated into Quick-CCA
OCR Engine (Devanagari & Latin Script)
A fully offline OCR system integrated into Quick-CCA's cheque clearing pipeline, extracting Devanagari and Latin-script fields with complete data privacy.
Project Overview
Built a fully offline FastAPI-based OCR system that extracts cheque fields (account number, amount, payee, date) in Devanagari and Latin script with high accuracy, ensuring complete data privacy with no cloud dependency. Transformer models were used to auto-generate structured training labels, accelerating custom OCR model fine-tuning while keeping all production inference on-device.
Key Features
- •Devanagari and Latin-script field extraction (account number, amount, payee, date)
- •Transformer-assisted structured label generation for OCR fine-tuning
- •Preprocessing: noise removal, skew correction, contrast enhancement
- •Fully offline, on-device inference — no cloud dependency
Technology Stack
- •Python, FastAPI
- •OpenCV, Tesseract
- •YOLOv5, Hugging Face Transformers
- •Docker, Nginx