Computer vision · University of Central Florida
I'm a PhD student at the University of Central Florida, advised by Dr. Yu Tian and Dr. Mubarak Shah. I work on long-horizon video generation and video diffusion models, and I also study AI agents in medicine.
My earlier research focused on image restoration, particularly image deblurring. Before joining UCF, I was a research assistant at the Cognitive Agents and Interaction Lab at the University of Dhaka.
My research focuses on extending video diffusion models to generate longer videos while maintaining consistency over time. I'm particularly interested in how models retain and use past context as a video unfolds. Alongside this work, I study AI agents in medicine, including how they use tools and knowledge to support clinical tasks.
PhD student in Computer Science
BSc in Computer Science and Engineering
CGPA 3.72/4.00 · Top 10% of class
Advised by Dr. Yu Tian and Dr. Mubarak Shah
Cognitive Agents and Interaction Lab, University of Dhaka
Advised by Dr. Md Mosaddek Khan
Papers and preprints. * denotes equal contribution.
Selective spatial memory
arXiv · 2026 New
A training-free controller that continually reassesses which observations to keep in a fixed-capacity spatial memory. By preserving views with few substitutes, Keepsake improves long-horizon scene consistency across MemCam and WorldMem.
Action-conditioned aerial video
arXiv · 2026
A controllable aerial video generation method that follows fine-grained inertial controls. Action tokens and a frozen physics probe improve motion alignment and visual quality on AeroBench.
Event-centric frame selection
arXiv · 2026
A training-free frame selection method for long-video understanding. Local curvature in query-frame relevance directs a fixed frame budget toward salient events, improving efficiency while preserving accuracy.
Real-world blur / sharp pairs
DMLR · 2026
A smartphone image deblurring benchmark with 42,045 paired images at 1920 × 1080 resolution across 843 scenes. Built from 240 fps slow-motion video, it exposes the challenges of real-world motion blur.
Restoring detail across domains
ICAART · 2026
A dual-domain architecture combining Vision Transformers with an FFT-ReLU frequency module. Spatial attention captures local and global dependencies while frequency-domain sparsity suppresses blur artifacts.
Clinical agents, tools & knowledge
Preprint · 2025
A systematic review of more than 140 studies on LLM-based medical agents. Our taxonomy connects clinical applications, autonomy, tool use, and knowledge integration, alongside the challenges of human oversight and trustworthy deployment.
Sharp
BlurryWACV · 2025
An FFT-ReLU sparsity prior for estimating blur kernels from a single image. The method works across a wide range of image types, with competitive restoration quality and up to twice as fast inference.
End-to-end authentication
Smart Health · 2023
A secure authentication scheme for wireless body area networks that treats patients’ phones as semi-trusted. It enables efficient communication between medical sensors and healthcare providers while resisting network attacks.