Subhashree Radhakrishnan
I am a Member of Technical Staff at Hark, working on multimodal AI and agents that connect perception, reasoning, and action.
My current work spans vision-language models (VLMs), computer-use agents, reinforcement learning, and self-improving agent systems.
Across my work, I focus on designing and scaling vision-language, video-language, and agentic systems, with a strong emphasis on 0→1 development—identifying the right problem, building the system, and taking it toward real-world deployment.
Previously, I was a Tech Lead and Applied AI Researcher at NVIDIA, where I led initiatives including NVCLIP, a multimodal foundation model for vision-language understanding; LITA, a Language-Instructed Temporal Localization Assistant for video reasoning; and 3D-Layout-R1, for structured spatial reasoning in 3D environments. I also led efforts to build reasoning agents for interactive gaming environments. My earlier work includes DiscoBox, an approach to weakly supervised segmentation.
My research interests include multimodal reasoning, continual learning, spatial and scene understanding, and agentic AI systems that integrate perception with decision-making and improve through interaction.
Prior to NVIDIA, I completed my Master’s at Virginia Tech, where I worked on human-object interaction in videos under the guidance of Prof. Jia-Bin Huang.
Filed 25+ patents across multimodal learning, video understanding, and spatial AI.
Using Neural Networks to Perform Object Detection, Instance Segmentation, and Semantic Correspondence from Bounding Box Supervision
Determining Associations Between Objects and Persons Using Machine Learning Models
Training Object Detection Models Using Transfer Learning
Automatic Labeling and Segmentation Using Machine Learning Models
End-to-End Action Recognition in Intelligent Video Analysis and Edge Computing Systems
Class Agnostic Object Mask Generation
Language Instructed Temporal Localization in Videos
Language Instructed Temporal Localization in Video Using Image Tokens, Video Tokens, and/or Soft Cross Entropy Loss
Techniques for Implementing Multimodal Large Language Models with Mixtures of Vision Encoders
Auto-Labeling Systems and Applications for Open-Set and Out-of-Domain Segmentation
Object Segmentation Using Machine Learning for Autonomous Systems and Applications
Point-Level Supervision for Video Instance Segmentation