Varun Kotte — Staff ML Engineer, Adobe

I build retrieval and LLM systems that stay reliable in production.

Reliability of RAG, retrieval, and LLM serving: calibration, conformal prediction, and knowing when a model should say it doesn't know.

6Adobe products on the platform

~7×Lightroom search click-through

50+citations, lead paper

Highlights

Architect of the LLM-serving and retrieval platform behind AI and search in six Adobe products.
Shipped Lightroom semantic search to general availability; search click-through rose roughly seven-fold.
First-named inventor of six on a published Adobe patent, since cited as prior art by a later granted U.S. patent.
Sole-authored papers accepted at workshops of CVPR, ICML, ICLR, and CompLearn.

A shared platform that seven product teams adopted instead of building their own. It backs AI and search features across six Adobe products.

Named-entity recognition and resolution behind intent-based photo search.

Calibrated abstention, model cascades, and routing under load.

AAAI · SDU

Adobe's production RAG method (co-author). Cited 50+ times by independent groups.

CVPR · GRAIL-V

Sole-authored. A training-free probe that catches answers not grounded in the evidence.

preprint

Sole-authored. Keeps structured output valid when the same prompt behaves differently across models.