Rohit Talluri

Rohit Talluri is co-founder and chief executive of Primitive Labs, which builds behavioral intelligence for AI-native software. Its agents model specific customers so product teams can test how people might respond to designs, features, and user journeys before releasing them.

A University of Washington graduate, Talluri began at Amazon Web Services as an enterprise solutions architect and later led solutions for its global mergers-and-acquisitions advisory practice. His work on cloud strategy for acquisitions addressed how companies organize cloud adoption during complex business transitions.

He subsequently focused on the infrastructure needed to train and operate foundation models. At AWS, he co-authored technical guidance on training Llama 2 with AWS Trainium, monitoring Trainium and Inferentia workloads, and resilient large-scale training with SageMaker HyperPod. In 2025, he helped introduce Inception Labs’ Mercury diffusion models to Amazon Bedrock and SageMaker JumpStart.

Talluri’s conversation with Cartesia co-founder Arjun Desai at AI Engineer World’s Fair 2025 explored how speech recognition, language models, and speech synthesis interact under demanding latency constraints. He emphasized offering developers specialized foundation models suited to particular applications and raised questions about edge deployment and the future role of human voice creators.

Talluri later joined Amazon’s AGI Autonomy Lab, working on computer-use agents and helping launch Amazon Nova Act. In 2026, he founded Primitive Labs with Jean Farmer and Gabriel Fong, backed by a16z speedrun and other investors. His founding announcement invoked Amazon’s empty-chair tradition: keeping the customer present in product decisions even when nobody representing them is in the room.

• Human behavior as a software primitive: Give builders a working model of customer preferences while products are being designed, not only after they ship.
• Behavioral fidelity over generic intelligence: Judge customer agents by how closely their decisions reflect a particular audience’s context and preferences.
• Production-ready foundation-model infrastructure: Make advanced models operationally viable through accelerator monitoring, resilient distributed training, and application-specific deployment choices.