At Samsung Research Institute Bangalore, I developed an automated 3D reconstruction pipeline for an in-house dataset captured with a custom camera and unknown intrinsics. Replacing traditional SfM point clouds with DUST3R outputs improved reconstruction quality from PSNR 30 to 45 dB.
I also built a LangSplat + CLIP pipeline for open-vocabulary object detection and segmentation in 3D Gaussian Splatting scenes, where text queries return object-region masks. I authored a research paper and am improving the manuscript for resubmission.
For Samsung’s smart-glass work, I developed lightweight on-device models and compressed an activity detection model fivefold with minimal F1-score drop. I also optimized tap detection latency from two seconds to 200 milliseconds.
On the SmartThings Team, I built and deployed a carbon emissions prediction model now running in production, and developed time-series models for solar energy prediction. My project work includes Graph Neural Networks for fracture simulation and an OpenGL rendering pipeline for Chandrayaan-II mission visualization.
