Jiostar chief architect details scaling AI discovery for 500m+ users using voice interfaces

Vijay Seshadri, Chief Architect at JioStar, discussed the challenges of deploying production-ready AI at massive scale, particularly in entertainment and live-sports platforms used by over 500 million consumers. JioStar integrated a conversational assistant into JioHotstar, allowing users to find content or cricket statistics via natural voice commands, supporting over 15 languages in approximately eight weeks. Seshadri noted that the system required balancing accuracy, latency, and cost, especially since over 60% of users preferred voice input. The development process involved rigorous evaluation, utilizing a maker-checker pattern where one LLM generated answers and another acted as a validator. Furthermore, domain expertise, such as input from sports-magazine editors, was crucial for developing accurate sports-statistics evaluations. Speaking at OpenAI’s Executive Summit in Singapore, Seshadri also advised leaders on AI deployment, stressing the need to monitor model efficacy in production beyond traditional infrastructure metrics. He emphasized focusing on measurable outcomes rather than just activity, and highlighted that human judgment is essential for safely deploying powerful tools like Codex.