Portrait of Kavyansh Chourasia

Kavyansh Chourasia

Research SDE 2

About

I am a Research SDE 2 at Microsoft Research India, where I build AI and cloud technologies designed to empower the Global Majority. I am passionate about addressing real-world challenges through inclusive innovation, ensuring that the benefits of the fourth industrial revolution reach diverse linguistic and cultural contexts worldwide.

Some of the projects I am currently driving or have previously led include:

  • Project Gecko & MMCTAgent: As a core contributor to Project Gecko, I focus on developing equitable generative AI for underserved regions. My work centers on MMCTAgent, a Multi-modal Critical Thinking Agent that uses a Planner-Critic architecture to enable sophisticated reasoning over massive collections of images and long-form video, bridging the gap between perception and deliberation in multimodal tasks.

  • VeLLM & Shiksha Copilot: Within Project VeLLM (uniVersal Empowerment with LLMs), I designed and led the implementation of Shiksha Copilot. This AI assistant empowers educators to generate personalized, culturally relevant lesson plans instantly. It is currently live in two states in India and has become a preferred tool for teachers looking to streamline their classroom preparation.

  • SEEDS (Scalable Educational Experiences with Digital Scaffolding): I led the system design and implementation of the SEEDS platform, which provides accessible, multimodal education for children with vision impairments. The platform supports diverse learning paths, including interactive self-learning through IVR calls.

  • Azure Purview: I contributed to this unified data governance solution by developing scalable methods for mining semantically sound data quality rules across multicloud and SaaS sources, ensuring rigorous adherence to customer privacy and data policies.

Currently, I am focused on advancing the frontiers of Multimodal LLMs. My research centers on improving video understanding—specifically optimizing how models ingest, reason across, and query hour-long video content to provide cited, context-aware answers in real-time.