Computer Science PhD Candidate University of Rochester
Google PhD Fellow Health Research, 2023–2026
Assistant Professor (on study leave), Department of CSE, BUET
I build machine learning and large language models to advance accessible healthcare -- focusing on digital biomarker extraction, video-based remote assessment, and LLM-driven clinical triage. Beyond research and teaching, I enjoy reading, traveling, and playing table tennis.
Over 900,000 people in the U.S. live with Parkinson's disease, yet many -- especially those over 65 -- struggle to access neurologists for diagnosis and monitoring. PARK (Parkinson's Analysis with Remote Kinetic-tasks) offers an accessible telemedicine solution. Using only a webcam, it guides users through motor, speech, and facial tasks and analyzes performance with AI. Through various studies, we show that PARK can screen for PD and track symptom severity with accuracy comparable to expert neurologists. Try a live demo here.
I completed two internships with the Digital Health Team at Samsung Research America, where I developed algorithms for heart-rate and non-invasive blood pressure monitoring using next-generation earbuds with IMU and PPG sensors. My contributions included analyzing BCG signals, building machine learning models for automated signal quality assessment, integrating these models into downstream applications, and designing study protocols for data collection and validation.
As first author, I developed KnowUREnvironment, a large-scale knowledge graph on climate change and environmental issues, automatically extracting over 210K domain-specific concepts and 411K relationships from scientific literature to support tasks like information retrieval, QA, and recommendation. I also designed SEER (Sustainable E-commerce with Environmental-impact Rating), a concept for E-commerce platforms that displays products' environmental impact. In a user study, SEER significantly increased eco-friendly purchasing behavior and achieved high usability scores, showing its potential to reduce carbon emissions from online shopping.
I co-developed XL-Sum, a large-scale multilingual summarization dataset with 1M article-summary pairs across 44 languages, including Bangla, enabling state-of-the-art results for low-resource summarization. I also helped create BanglaBERT, the first Bangla-specific BERT model, along with new benchmarks and datasets, establishing strong performance across multiple language understanding tasks.
I introduced KICQ (Keyword-aware Influential Community Query), a framework for finding influential communities in large attributed graphs. KICQ combines semantic keyword matching with a novel influence measure that accounts for both cohesiveness and dominance, enabling more intuitive and effective community search. We further designed efficient algorithms to scale this approach and validated its performance through experiments and case studies.
The work explored the use of smartwatches and fitness trackers for real-time cardiovascular monitoring, with a focus on Atrial Fibrillation (AF) detection from noisy photoplethysmography (PPG) signals. Our team developed BayesBeat, a Bayesian deep learning model that not only improves AF risk prediction but also provides uncertainty estimates. BayesBeat outperforms existing methods while being 40-200x more parameter-efficient, making it practical for deployment on resource-constrained wearable devices. As a collaborator, I played a key role in preparing and presenting the grant proposal that secured 1.5 million BDT funding from the ICT Division, Bangladesh for this project.
We developed the first supervised learning-based tool for detecting NoSQL injections, supported by a novel training dataset that I specifically designed for this task. The tool achieved a 0.93 F2-score through cross-validation and outperformed the existing solution (Sqreen) by 36% in detection rate. It is also database-agnostic, demonstrating consistent performance across MongoDB and CouchDB.
Result processing of about 1.5 million applicants for admission into 7,447 colleges of Bangladesh. This was a government-funded project to automate the entire higher secondary admission system. I designed and implemented the core algorithms for result processing using JAVA and Oracle database.
A government-funded project for the office automation of ICT Division and Planning Commission of the Peoples' Republic of Bangladesh. I actively contributed as a technical consultant toward designing and developing the "meeting and event management" and "accounting" modules.
There are multiple customers and multiple workers in a system. A customer can order something online, then a worker needs to go to the task location from his current location, pick the order, then deliver this to the home address of the customer. Google map API was used to estimate the duration of such travels. The spatial organizer distributes tasks among the employees so that the throughput is nearly optimal and all the employees are fairly loaded. The project was an android application of which, I designed and implemented the back-end algorithms in PHP.
An android application developed for Dhaka City Corporation. The citizens can report road and transportation-related issues to authority using this app. This project was initially developed for BRACathon 2015, a software development contest. After being the champion, this project was funded by BRAC. I designed and implemented the back-end using the PHP Codeigniter framework.
Fantasy cricket is a game where a user forms a virtual team consisting of the players from the real game and receives points based on the performance of those players in the actual game. I designed the database and implemented the back-end using the PHP Codeigniter framework.
I am an avid reader and I love to collect quotes. Here are some of my favorites.