Case Study
UrduOCRTextExtractionPipeline
Digitizing legacy legal and public records in Pakistan. Built an extraction pipeline using Qwen3 VLM and a fine-tuned ByT5 model to fix looping artifacts and improve accuracy.
The Challenge
Traditional OCR failed on cursive, nastaliq Urdu scripts, and early VLM approaches introduced severe looping artifacts when processing low-quality scanned legal documents.
The Solution
Implemented a two-stage pipeline: a Qwen3 VLM for initial structural layout parsing, followed by a ByT5 model fine-tuned on a proprietary dataset of 500,000 verified document pairs to eliminate token looping.
The Results
- Achieved 96.5% character accuracy rate (CER) on degraded documents.
- Processed 2.5 million historical records in the first month of deployment.
- Reduced manual verification time by 85%.