Researcher · Founder · Ph.D., IIT Delhi
Learning to see, understand,
and build with AI.
I work on computer vision and multimodal learning: how models represent the world, generate visual content, and become useful AI systems.
I founded OmniTensorLabs, where I work on multimodal systems, agents, and evaluation. I also advise Timble AI, where I previously served as CTO.
Previously, I was AI Architect / Director of ML at Typeface. I co-founded TensorTour as CTO; Typeface acquired the company in 2024.
At C-DAC, I led national electoral search across 900M+ voter records and provided technical leadership for e-RaktKosh, serving 2,100+ blood banks by 2021.
1,171 citations 18 h-index 22 i10-index
Google Scholar · All-time metrics ·
Selected research
Three threads through my work: efficient 3D representations, controllable generation, and learning across modalities.
3D vision / Efficient foundation models
Distilling 3D Foundation Models for the Edge
How much of a 3D foundation model can a smaller model retain? Foundry compresses rich point-cloud representations into compact tokens, preserving their usefulness across downstream tasks.
Generative AI / Object preservation
Controllable Image Synthesis with Object Preservation
Changing a scene should not change the object that matters. This work studies controllable image synthesis that preserves an object while adapting its surroundings, composition, and lighting.
Multimodal learning / Shared representations
A Novel Transformer Based Network for Large Scale Multimodal and Multitask Learning
Can different modalities benefit from learning together? OmniVec2 combines modality-specific input processing with a shared transformer and task-specific heads, extending the cross-modal learning explored in OmniVec.
For the full publication list and citations, visit Google Scholar.
Current work
Alongside publications, I build tools for developing AI systems, evaluating their behavior, and understanding the research landscape.
Research & products
My lab’s work on visual, multimodal, and agentic intelligence, connecting research questions with working demonstrations and AI products.
Explore the labOpen-source evaluation
A Python library for evaluating AI systems. Bringing repeatable evaluation into the process of building models and agents.
Explore MultivonResearch exploration
An interactive view of topic trends across major machine learning conferences, built to make the literature easier to navigate.
Open dashboardFrom the notebook
I’m happy to hear from researchers, engineers, founders, and anyone interested in the work.