At National Institute of Technology Surat’s Computer Vision Laboratory, I study how Vision Transformer representations evolve across depth and training. I developed the Transformer Geometry Observatory, a research program with checkpoint-wise PyTorch pipelines and quantitative analyses of neural representations.
I also developed COAST, a dynamic token-routing framework for adaptive computation in Vision Transformers. Previously, as an Advanced Vision Research Intern at National Institute of Technology Surat, I developed AvioDepth, a lightweight sequential-attention framework for real-time monocular depth estimation in UAVs, resulting in a peer-reviewed conference publication.
