Umar Hasan - Homepage

About

I am a Computer Science undergraduate at North South University with research experience in artificial intelligence, machine learning, and healthcare applications. My work spans multiple domains including trustworthy AI systems, medical foundation models, and explainable deep learning for clinical diagnostics.

My current research interests include developing safety mechanisms for multi-agent LLM systems, creating culturally-grounded benchmarks for underrepresented linguistic communities, and applying computational methods to analyze public discourse on AI. I am particularly focused on building AI systems that are trustworthy, interpretable, and aligned with diverse human values.

Currently, I am a research intern at Duke University's Kamaleswaran Lab, contributing to research on medical foundation models for physiological time-series forecasting. My work involves developing multimodal architectures that fuse ECG and PPG signals for comprehensive physiological representation.

Research Experience

Research Intern

Kamaleswaran Lab, Duke University | Aug 2025 – Present | Remote

Co-authoring a comprehensive review on medical foundation models for time-series forecasting, responsible for synthesizing and analyzing parameter-efficient fine-tuning (PEFT) and domain adaptation methods
Contributed to the ongoing development of a novel multimodal foundation model that fuses ECG and PPG signals for comprehensive physiological representation and conversational AI-based querying
Implemented a Neural Ordinary Differential Equation (Neural ODE) Transformer as a core component of the model's feature fusion architecture

Research Intern

Machine Intelligence Lab, North South University | Sep 2025 – Present | Dhaka, Bangladesh

Contributing to research projects in applied computer vision for engagement detection and quantum machine learning (QML)
Developing deep learning architectures to model temporal features from facial expressions, eye gaze, and head pose for engagement classification on benchmark datasets (DAISEE, CMOSE)
Investigating foundational QML methods, including the encoding of classical data into qubits and the training of parameterized quantum circuits

Research Collaborator

Jouf University | Oct 2025 – Present | Remote

Co-authoring a systematic literature review, "Deep Learning in Knee Osteoarthritis," with an assistant professor from Al-Jouf
Responsible for creating the original manuscript draft, synthesizing research on model architectures and XAI, and developing data visualizations
Collaborating on an intensive, deadline-driven research project targeting a high-impact journal submission

Publications

Workshop Papers

Umar Hasan
"BIICK-Bench: A Bengali Benchmark for Introductory Islamic Creed Knowledge in Large Language Models"
Accepted at the 5th Muslims in ML Workshop co-located with the 39th Conference on Neural Information Processing Systems (NeurIPS 2025)

[Accepted Manuscript PDF]

Conference Proceedings

Umar Hasan, Mohammad Abdul Qayum
"Predicting and Explaining Fatal Road Casualty Types in Great Britain: A Comparative Analysis of Machine Learning, Deep Learning, and Transformers"
To Appear in S. Palaiahnakote et al. (eds), Data Science, AI and Applications: First International Conference, ICDSAIA 2025, Dhaka, Bangladesh, July 18–19, 2025, Proceedings, Part I. Communications in Computer and Information Science. Springer, Cham.
Invited for journal extension to Springer-Nature Computer Science (Q2)

[Accepted Manuscript PDF]

Ibrahim Sani*, Umar Hasan*, Raihan Sharif, Tahfeem Islam Siam, Md Alamgir Hossain, and Riasat Khan (*Equal Contribution)
"Data-Driven Forecasting of Refugee Displacement Using Machine and Deep Learning Models with XAI"
In Press for the Proceedings of the ACS/IEEE 22nd International Conference on Computer Systems and Applications (AICCSA 2025), IEEE Xplore

[Accepted Manuscript PDF]

Manuscripts Under Review & In Preparation

Umar Hasan, Muhammad Rafsan Kabir, Abdullah Al Raiyan, Md. Sifat Haque Zidan, and Sifat Momen
"Distilling Ensemble Knowledge via a Teacher-Assistant for Explainable Cervical Cancer Screening on Edge Devices"
Under Review at PLOS Digital Health (Q1, IF: 7.7)
Proposes AgileNet, an ultra-lightweight (0.29M parameters) student model for cervical cancer screening, achieving 98.41% accuracy through a Teacher-Assistant Knowledge Distillation framework.

Md Hassanuzzaman, Umar Hasan, Tilendra Choudhary, and Rishikesan Kamaleswaran
"Medical Foundation Models for Time Series Forecasting: A Comprehensive Review of Architectures, Methodologies, and Challenges"
In Preparation for submission to Artificial Intelligence Review, Springer Nature (Q1, IF: 13.9)

Mohammad Mohiuddin Azad, Umar Hasan, and Zarif Muhtasim Showgat
"Deep Learning in Knee Osteoarthritis: A Comprehensive Review of Imaging Modalities, Model Architectures, Grading Systems, and Explainable AI"
In Preparation for submission to Artificial Intelligence in Medicine, Elsevier (Q1, IF: 6.2)

Umar Hasan, Riasat Khan
"CerebroNet: Ultra-Lightweight Brain Tumor Classification via Knowledge Distillation and Attention Optimization on a South Asian Cohort"
In Preparation for submission to Brain Informatics, SpringerOpen, Springer Nature (Q1, IF: 4.5)
Introduces CerebroNet, an ultra-lightweight (0.637M parameters) CNN achieving 95.61% accuracy, validated on a novel South Asian cohort.

Ongoing Research Projects

Mitigating Propagation of Unreliable Information in Multi-Agent Large Language Model Systems
Developing and evaluating strategies to identify and mitigate propagation of unreliable information in collaborative LLM agents.

Mapping AI Discourse on Reddit: A Mixed-Methods Analysis of Dominant Themes and Narratives
Applying BERTopic and thematic analysis to understand public perceptions and narratives surrounding AI on platforms like Reddit.

Teaching Experience

Undergraduate Teaching Assistant

North South University | Jul 2024 – Dec 2024 | Dhaka, Bangladesh

Mentored and supported 152 undergraduate students across Calculus I & II and Business Mathematics courses
Graded and provided constructive feedback on assignments and quizzes to enhance student understanding

Professional Service

Peer Reviewer | Oct 2025
The 5th Muslims in ML Workshop at NeurIPS 2025
Reviewed two anonymized manuscript submissions via OpenReview on the topics of ML/DL applications

Memberships
Student Member, Institute of Electrical and Electronics Engineers (IEEE)

Selected Projects

Et Al.: A Web-Based Collaborative LaTeX IDE

Jan 2025 – Jun 2025 | Frontend Developer

Built the entire client-side application using Vue 3 and TypeScript, including a CodeMirror 6 editor with LaTeX syntax highlighting
Implemented real-time collaborative editing functionality by integrating the Yjs framework (CRDTs) and a Socket.IO client to synchronize changes and user cursors

Tech Stack: Vue.js, TypeScript, Yjs, Socket.IO, Node.js, MongoDB

Technical Skills

Languages: Python, Java, TypeScript, JavaScript, C++, C, SQL

Frameworks & Libraries: PyTorch, Scikit-learn, Pandas, NumPy, Vue.js, Node.js

Tools & Technologies: LaTeX, Git, Docker, MongoDB, Linux, REST APIs

Languages

English: IELTS Overall Band Score 8

Bengali: Native Proficiency

Arabic: Elementary Proficiency