Multi-Modal Document Understanding for Low-Resource Languages

June 2025 - Present | Independent Collaboration

Supervisor(s): Dr. Md Mofijul Islam, Dr. AKM Mahbubur Rahman, Dr. Aman Chadha

• Collecting, annotating, and curating government forms in South Asian low-resource languages for document understanding tasks such as key information extraction, classification, and entity linking.
• Building a synthetic document data generation pipeline to address low-resource scarcity and ensure demographic and regional representativeness.

Synthetic Data Based Agentic Visual Question Answering (VQA)

Link, May 2024 - Present | Independent Collaboration

Supervisor(s): Dr. AKM Mahbubur Rahma, Dr. Md Mofijul Islam, Dr. Aman Chadha

• Working on building large-scale synthetic question-answer-rationale dataset for interpretable VQA tasks.
• Applying zero-shot, one-shot, and few-shot based CoT (Chain of Thoughts)-style dynamic prompting for diversification of synthetic question-answering data.

Textual Context Integration for Indoor Sign Detection

Link, Jan 2023 - May 2023 | Wichita State University

Supervisor(s): Dr. Vinod Namboodiri

• Developed multi-modal fusion-based transformer models to detect indoor sign for BVI (Blind and Visually Impaired).
• Prototyped a multi-modal chat interface for cooperative guided movement of low/no-vision people in indoor settings.

Human in the Loop ML in Computational Healthcare

Link, Jan 2019 - Dec 2019 | University of Dhaka

Supervisor(s): Dr. Md Mofijul Islam, Md Mahmudur Rahman

• Investigated interpretability and domain-awareness in pulmonary disease detection using CNNs.
• Observed joint effects of trainable filters and customized domain-specific non-trainable filters on CNNs disease detection capability.