Aarush Selvan is a product manager at OpenAI working on search for ChatGPT. At Google, he originated Gemini Deep Research, an AI research assistant that investigates complex questions across the web and produces source-linked reports, and the Gemini Extensions platform, which connected the assistant to other applications and services.
Originally from the United Kingdom, Selvan studied symbolic systems at Stanford and began his career at McKinsey & Company. At Google, he worked on AI infrastructure, including specialized chips, compilers, and machine-learning frameworks, as well as speech and language technology for Google Assistant. His published work examined large-model training with Cloud TPU v4 Pods and Google’s MLPerf benchmark results.
Gemini Deep Research, launched in December 2024, addressed a specific weakness of conventional chatbots: they often outline how to research a question without resolving its consequential details. Selvan helped shape a product that drafts editable research plans, browses sources, incorporates new findings, and produces cited reports. His account of research-agent design emphasizes several principles:
- Research plans users can steer: Showing an editable plan before execution lets people correct the assistant’s priorities; visible sources make its conclusions easier to inspect.
- Asynchronous research inside conversational products: Multi-minute investigations require clear progress, practical handling of long reports, and interfaces that support follow-up questions.
- Inference-time compute for better answers: Additional browsing and reasoning are valuable when they resolve concrete uncertainties that immediate responses leave unanswered.
- Personalized strategic synthesis: Effective research should adapt its sources, analysis, and deliverables to the user’s profession and purpose, potentially incorporating coding, scientific analysis, or financial modeling.
At AI Engineer Summit 2025, Selvan described moving beyond information aggregation toward assistants that identify implications and produce analysis tailored to individual users.
