I'm a Member of Technical Staff at Stellon Labs, where I work on text-to-speech (TTS) and automatic speech recognition (ASR) models. I joined in July 2026.
Previously, I was a research fellow at Sankara Eye Hospital and Microsoft Research India, where I worked with Dr. Mohit Jain, Dr. Vineeth N B, and Prof. Thomas Schultz.
My research spans speech, vision-language models, and healthcare AI. My previous work includes task-transfer dynamics in vision-language models, surgical scene understanding in cataract surgery, and expert-in-the-loop LLM chatbots for patients and community health workers.
I received my B.Tech. in Computer Science from Delhi Technological University (DTU) in 2023. During my undergrad, I interned at Prime Video, Amazon and Kroop AI.
We introduce a parameter-efficient adaptation of BAGEL using Group Relative Policy Optimization (GRPO) with Low-Rank Adaptation (LoRA) for interleaved visual reasoning. BAGEL-RL improves average benchmark performance from 62.49% to 70.36% while updating only 2.6% of the model parameters.
We investigate how finetuning on one task affects zero-shot performance on others across the Qwen-2.5-VL model family. We introduce the Perfection Gap Factor (PGF) to characterize task-to-task transfer and leverage task interactions to identify effective proxy datasets.
A training-free, lightweight frame selection algorithm for long-video understanding. We model a video as a differentiable trajectory in visual latent space and use a Taylor expansion to predict future features; frames with large residuals ("temporal surprises") are selected as keyframes. Adds only 0.02× overhead while improving VQA accuracy by up to +12.5 points.
We present the first comprehensive dataset for MSICS cataract surgery tool segmentation and introduce ToolSeg, a phase-informed segmentation method that leverages surgical phases to enhance tool segmentation accuracy.
We developed CataractBot, an LLM-powered chatbot answering cataract surgery questions via a curated knowledge base with expert-verified responses, supporting multimodal and multilingual inputs.