Georgia Institute of Technology
Implemented a stitched LLM architecture using separate models for prefill and decode phases, improving both TTFT and TBT while maintaining model accuracy. Achieved 5% decrease in latency than the baseline model (using same bigger model in both the phases). Tested out generalization capability for the architecture on different tasks (q/a, summarization, code completion, math problems)
Georgia Institute of Technology
Built a PyTorch-based point-wise neural network with physics-informed loss constraints that maps spatial coordinates and design parameters to aerodynamic flow fields. Transformed a data-limited problem (500 configurations) into 35M training samples through pointwise prediction, achieving fast surrogate modeling for CFD simulations with R² > 0.95.
Georgia Institute of Technology
Developed a multiplayer game using Unity and C# where players must survive against waves of zombies. Implemented features like player customization, weapon upgrades, and a dynamic environment that changes based on player actions. Check out the trailer below
Smart India Hackathon
Developed an end-to-end pipeline to find relevant books/journals from the database given 4-5 keywords. Winner solution at Smart India Hackathon 2019 and the solution is set up in PD Hinduja Hospital and Research Center.
Indian Institute of Information Technology Kalyani
Proposed an end-to-end approach for correcting errors in annotation-affected document images. The method removes annotations using advanced computer vision techniques and applies a spelling correction model to OCR-extracted sentences, achieving a BLEU score of 10.64.