Associate Data Scientist
Lahore, Punjab, Pakistan
Led the enhancement of OCR models, achieving performance comparable to Google OCR. Focused onoptimizing text detection and recognition across a diverse set of documents. Tried multiple models including PaddleOCR, Tesseract, EasyOCR, Deep Text Recognition Benchmark, TrOCR, ABINet, YOLO, andCRAFT-pytorch, and achieved state-of-the-art results by combining PaddleOCR and YOLO. Successfullyimproved accuracy and efficiency in recognition and detection tasks.Leveraged state-of-the-art LLM models, including LLama3 (8B), Qwen2, ChatQA-1.5-8B, LLama2 (7B),Mistral (7B), PHI-2, RoBERTa-XLM, BERT, and Flan-T5, to tackle complex information extraction tasksfrom OCR across diverse languages and document types. After extensive evaluation, finalized LLama3 and Qwen2 for their superior performance. Successfully extracted essential data from a variety of documents, achieving performance comparable to the Google PALM model.