Saroj Kumar Behera
Title of the Talk: Your Prompts Are Your Preferences: Building Recommendations from LLM Queries
Abstract:
People tell AI assistants things they’d never type into a search box. They describe their goals, their constraints, and what they’re stuck on, often in full sentences. That makes a prompt history one of the richest preference signals we’ve ever had, and most products throw it away after each session. This talk walks through how to turn LLM queries into a working recommendation engine. We’ll cover how to embed and cluster query histories into interest profiles, how to separate what someone wants right now from what they care about over months, and how to retrieve relevant items from millions of candidates using approximate nearest neighbor search. We’ll also look at the problems that show up in practice: interest drift, noisy one-off prompts, the curse of dimensionality in user embeddings, and the privacy line between helpful and creepy. You’ll leave with a reference architecture and a clear sense of where this approach works and where it doesn’t.
Bio
Saroj Kumar Behera is a Senior Engineering Manager with more than 12 years of experience building large-scale data and software platforms, most recently at leading fintech companies. He leads engineering teams working on data engineering, cloud platforms, and AI-driven systems, with hands-on depth in AWS and Snowflake. His current interests include high-dimensional retrieval, embedding-based matching, and practical ways to turn LLM interactions into useful product signals. Saroj holds an M.S. in Computer Science from Texas Tech University.

