About me
I am a Ph.D. candidate working with Dr. Ali Jannesari at Iowa State University.
My current research explores algorithmic and systems-level optimizations for GNNs, aiming to reduce computational and communication overhead while maintaining model accuracy.
I am particularly interested in efficient data movement, parallel computing, memory management, and distributed optimization tailored for modern HPC platforms.
Recent Updates
Projects
-
NOMAD
A distributed-memory graph embedding framework using MPI to generate embeddings for massive graphs. NOMAD demonstrates median speedups of 10/100x on CPU-based NERSC Perlmutter cluster relative to the popular reference implementations of multi-threaded LINE and node2vec, 35-76x over distributed PyTorch-BigGraph (PBG), and competitive embedding quality relative to LINE, node2vec, and GraphVite, while yielding 12-370x end-to-end speedups on real-world graphs. [arXiv]
-
Rudder
LLM agents that adaptively steer prefetching and buffer management in AWS DistDGL to reduce communication overhead in distributed GNN training. Evaluated on NERSC's Perlmutter Supercomputer, Rudder improves end-to-end training performance by up to 91% over baseline DistDGL and reduces communication by over 50%. [Code] [ICS'26]
-
MassiveGNN
A novel prefetch and eviction scheme integrated into Amazon's DistDGL to optimize sampling and communication overheads in distributed Graph Neural Network (GNN) training, achieving a 15-40% performance improvement in Open Graph Benchmark (OGB) graph datasets on NERSC's Perlmutter Supercomputer. [Code] [CLUSTER'24]
-
Model Souping in GNNs
Souping strategies for GNNs that enhance efficiency by reducing time and memory overhead. Our methods achieve significant speedups and accuracy improvements across various datasets and architectures. [Code] [IPDPS'25 GrAPL]
-
Scalable Spatiotemporal Forecasting for Hydrology
Hydrological forecasting using CNNs, LSTMs, GNNs, and Transformers, integrating domain knowledge to improve accuracy, efficiency, and transferability on large-scale river basins, with distributed training for scalable deployment. [NeurIPS'21 AI4S] [HPC ASIA'23 MMCP] [IPDPS'23 ExSAIS] [SIGSPATIAL'25]
Patents
-
TowFi
A mobile network dead zone warning system, deployed on the IBM Bluemix, which tracks and warns users about network depleted zones in the users’ vicinity using a combination of statistical analysis and interpolation of previously available networks data. Won IBM Bluemix Idea Hack in association with NASSCOM & IAMAI Mobile. Patent Filed: 201731014184 (April 2017, India)
-
SnaaP Tag
A community-based low-cost hybrid energy sensor network system using RFIDs for real-time localization and tracking of objects. Received funding from the Department of Science & Technology, Govt. of India. Patent Filed: 201731014175 (April 2017, India). [Code]
Course Projects
-
Product Recommendation System
A system that recommends Amazon products to users based on product ratings, only these ratings are predicted by the model itself by analyzing the sentiment from the users’ comments.
-
Smart Toilet
A Raspberry Pi-based system designed to integrate into the toilet capable of detecting blood in urine and stool. It is paired with an Android Application that sends alerts if an abnormal amount of blood is detected. This system was selected for presentation at a senior care facility to showcase its potential for early health monitoring.
-
Cerca Trova
A real-time rapid response emergency redressal service where an Android Application is designed to bridge the gap between the help seeker and the emergency service provider. [Code]