Anoop Kotha works in applied AI at OpenAI, helping customers build dependable products with advanced language models. He coauthored the GPT-5 prompting guide and develops practical approaches to production voice agents, model delegation, and evaluation.
Raised in Michigan, Kotha attended the University of Michigan and worked as an engineer at Retool before joining OpenAI. His customer-facing work addresses the product decisions that shape real-world AI systems: response speed, conversational quality, tool access, safety, and reliability. At AI Engineer World’s Fair 2025, he and Toki Sherbakov outlined an architecture for building effective voice agents.
- GPT-5 prompting: His coauthored guidance shows developers how to calibrate agentic eagerness, reasoning effort, response verbosity, and instruction clarity when building coding applications and autonomous workflows.
- Specialized-model delegation: A realtime voice model handles immediate conversation while stronger reasoning models resolve consequential tasks, such as interpreting return policies. Restricting each agent’s tools and passing summarized conversation state between agents improves control without sacrificing continuity.
- Voice-agent evaluation: He recommends starting with observability and expert-labeled examples, then combining transcript-based checks of business rules and tool use with audio-specific assessments of pacing, tone, and delivery. Simulated conversations with synthetic customer personas can expose additional failure modes.
- Asynchronous safety guardrails: Because generated text can arrive before its corresponding speech finishes playing, safety checks can run during that interval, preserving responsiveness while screening the interaction.
