JioStar builds AI discovery assistant for 500M-plus viewers

One interface for entertainment and cricket

The assistant lets viewers speak naturally to find movies, shows, historical highlights, live-match moments and cricket statistics. Results appear through a visual, conversational experience rather than a conventional search interface.

JioStar Chief Architect Vijay Seshadri described the product as a “voice-visual interface” in a conversation with Sidharth Sharma, OpenAI’s Head of Go-to-Market for ASEAN, at OpenAI’s Executive Summit in Singapore. The discussion was part of OpenAI’s Executive Function series, which features leaders working on AI transformation.

The system was designed for two distinct types of discovery. Entertainment viewers might ask for something that fits their mood, request a surprise recommendation or describe a film without knowing its title. Cricket fans might search for a fall of wickets, a series of sixes, a match linked to a player’s score or statistics about every ball.

Speed, accuracy and cost shaped the design

JioStar operates entertainment and live-sports platforms used by more than 500 million consumers. Its platforms have supported as many as 72.5 million concurrent live-sports viewers, making production performance a central consideration for the project.

Seshadri said the team framed its technical choices around a three-way tradeoff among accuracy, latency and cost. Voice made that balance especially important because more than 60% of JioStar’s users use voice rather than text. The assistant therefore had to return relevant answers quickly while remaining economically viable at scale.

The work began in October and November, ahead of a major cricket tournament in February. JioStar worked with a small group of data scientists and engineers, alongside OpenAI’s forward-deployed engineers, to identify the capabilities that had to work reliably.

Guardrails and evaluation came before expansion

The team first decided what the assistant should not answer, because an open voice or text interface can invite almost any question. JioStar then created a seed dataset for evaluations and used models to expand it with more variations.

The company used a maker-checker approach in which one model helped determine how to answer while another served as a validator. Seshadri said designing that validator created a further evaluation challenge: “who tests the judge?” Domain experts, including sports-magazine editors, helped define realistic questions for cricket-statistics testing.

OpenAI’s Codex helped accelerate tasks such as generating evaluation cases and updating the system. Seshadri stressed that faster development did not eliminate the need for human judgment, validation and production guardrails. He also said he would invest earlier in monitoring model efficacy, alongside traditional infrastructure measures such as P99 latency.

What the project means for AI deployment

JioStar’s account presents AI discovery as a production engineering problem as much as a model-selection exercise. The project combined multilingual interaction, voice input, entertainment recommendations and detailed sports data under tight time and scale constraints.

For enterprises moving beyond experiments, Seshadri advised focusing on outcomes rather than visible activity. He said generated lines of code do not show whether meaningful value was created, and argued that production AI requires people who can apply tools with judgment. In his view, strong engineers can multiply their impact with tools such as Codex, but weak validation can multiply risk at the same time.