Parag Gawai

Fine-tuning, Training, and Evaluating Donut Multimodal Transformer Model. Prepared the image dataset from pdfs and tiffs using Hugging Face Build and Load in Apache Arrow format with Multiprocessing. Fine-tuned and Trained Donut Transformer for classification task with different tokenization technique for better performance. Achieved 85% accuracy and 0.41 F1 score, analyzed errors, compared with other text models. Text Extraction from documents using Different OCRs. Leveraged different types of OCRs, including Google Cloud Vision OCR and AWS Textract, to extract text from the documents. Utilized GPT-4o model to check whether the document is OSV or not. Achieved accuracy of 93% with exact match using GPT-4o.

Serving as Teaching Assistant for Data Analytics Lab of CORAL Department, IIT Kharagpur by teaching around 25 students.

Developed a model for Satellite and Model oceanic current Data Validation with in-situ drifter data from eight large datasets. Worked on Linux Server associated with High Performance Computer to calculate Correlation Coefficient and RMSE. Performed the yearly and monthly analysis for the year 2020 for Indian Ocean, Arabian Sea, Bay of Bengal and Equatorial IO.