About

I'm Varun Kotte, a Staff Machine Learning Engineer at Adobe in San Jose. I build the retrieval and LLM systems that decide when a model's answer can be trusted before it reaches a user.

At Adobe I build the backend LLM-serving and retrieval infrastructure behind AI and search features in six products, including Acrobat, Lightroom, Photoshop, and Express. My work spans the parts of the stack most people don't see: the serving platform, the named-entity recognition behind semantic search, and the reliability layer that decides when a model should be trusted and when it should not.

My focus is calibration, conformal prediction, retrieval that does not hallucinate, and systems that fail honestly. Outside of Adobe I write and publish on these questions, and I'm a named inventor on a published Adobe patent application for closed-domain question answering. The hard part isn't getting a model to answer; it's knowing when its answer shouldn't ship. That boundary is what I build and what I publish on.