Sriram Selvam

Copilot Search

Active

Copilot Search seamlessly blends the best of traditional and generative search together to help you find what you need.

I work on identifying and integrating enrichments—such as answer cards, videos, and images—with LLM-generated text responses. Along with my team, I build and maintain the online enrichment engine that serves millions of users every day.

Generative Search

Active

This experience combines the foundation of Bing's search results with the power of large and small language models (LLMs and SLMs). It understands the search query, reviews millions of sources of information, dynamically matches content, and generates search results in a new AI-generated layout to fulfill the intent of the user's query more effectively.

I'm one of the core founding team members behind Bing's Generative Search result generation framework (patent pending). Together with my team, I designed and built the initial framework for enrichment data retrieval, LLM/RAG-based enrichment pairing techniques, and LLM-driven evaluation and metrics. This pipeline has powered the generation of millions of enriched result pages, significantly enhancing user experience at scale.

PANORAMA: A synthetic PII-laced dataset for studying sensitive data memorization in LLMs

Active

I co-developed PANORAMA, a large-scale synthetic dataset of 384K samples from 9.6K realistic human profiles to model the distribution and context of PII in online content. This was developed specifically to help researchers working in the space to reliably evaluate their mitigation strategies and quantify various modeling techniques for their efficacy at not remembering sensitive private information.

Using OpenAI's o3-mini, we generated diverse web-native formats—social posts, reviews, wikis, and more—with embedded sensitive data. By fine-tuning Mistral-7B across varying data replication rates, we analyzed memorization patterns across content types, providing insights into privacy risks in LLM training. PANORAMA enables robust model auditing and privacy-preserving research, with open-source data and tooling.

Creator and Code Assist

Past

This feature helps user get tailored LLM generated responses for their tasks with minimal prompting. Feature focuses on various areas such as email writing, story creation, code queries and creation.

I developed the prompts and the prompt templating engine that powers this experience. In addition, I was responsible for assembling the UX customization panel, enabling flexible and user-friendly interactions.