JioStar’s chief architect on AI discovery at 500M-user scale

Vijay Seshadri, Chief Architect at JioStar, discussed the company’s deployment of a conversational AI assistant on JioHotstar, allowing users to speak naturally to find movies, shows, live-match moments, or cricket statistics. The system returns results through a visual, conversational interface, addressing the limitations of traditional recommendation engines that rely on iterative suggestions rather than contextual understanding.

The project was developed in about eight weeks to support two distinct audiences: entertainment viewers and cricket fans, each with different discovery needs. Entertainment users might request a surprise recommendation, while sports fans could ask for specific match updates or detailed cricket statistics. The team balanced accuracy, latency, and cost, with over 60% of users preferring voice input for its convenience.

Early in the process, the team established guardrails to define what the system would not answer, preventing open-ended queries from derailing performance. They used a seed dataset and expanded it with model-generated variations to create a robust evaluation framework, incorporating a maker-checker pattern where one model generated answers and another validated them.

Seshadri emphasized the importance of production monitoring for model efficacy, noting that traditional infrastructure metrics like P99 latency are insufficient. He also highlighted the need to focus on outcomes rather than activity, cautioning against mistaking visible progress for meaningful value, especially when companies emphasize input metrics like lines of code generated.