Gaurav Joshi

I’m a Computer Science MS student at UC San Diego, working at the intersection of computer vision, graphics, and agentic AI.

Most of what I build falls under vision agents — agents that reason over the images they encounter while doing a task, instead of reasoning around them. My past research with Manmohan Chandraker had agents learn Blender material generation by watching YouTube tutorials, extracting procedural knowledge from how experts actually build materials to guide the agent. I’m now working on harness engineering for creative software, the plans, tools, and feedback loops that let an agent actually operate inside something like Blender. I write about what actually makes this work on my blog.

I also like taking on things I don’t know how to do yet: after undergrad I co-founded FluenTalk, a voice-based AI assistant for practicing spoken English, built end to end and tested with 200+ students. I shelved it once grad school started, but it left me wanting to build things people actually use and pay for, not just write papers about.

Before UCSD, I was at IIT Gandhinagar, working on inverse modeling for outdoor scenes — reconstructing and re-rendering them from a single image with Gaussian splatting and NeRFs. In between: a summer at Purdue’s IDEAS Lab on neural-operator surrogates for fluid simulation, and a Google Summer of Code cycle with OpenAstronomy/Stingray, writing a Gaussian-process tool to detect quasi-periodic oscillations in astronomical time series.

I like closing the loop between seeing and doing, in whatever form that takes.