JioStar’s voice-first content search shows what production AI needs at 500 million users

JioStar, which runs entertainment and live-sports platforms used by more than 500 million people and has supported as many as 72.5 million concurrent live-sports viewers, built a conversational assistant into JioHotstar using the OpenAI API. Viewers can ask for shows, match highlights or cricket statistics in natural language across more than 15 languages, and the team says it built it in about eight weeks ahead of a major cricket tournament. The conversation took place at OpenAI’s Executive Summit in Singapore, where OpenAI’s forward-deployed engineers were described as working alongside JioStar’s small team.

This is a promotional interview on OpenAI’s site with no outcome numbers like engagement or accuracy, so it doesn’t grade the product. What’s worth attention is how the team got to production, where most enterprise AI projects stall, the same lesson from Gilbert + Tobin’s rollout (coverage).

The team decided first what the system should not answer. For an open voice interface used by hundreds of millions, that is the right order, and the opposite of how many demos get built. Media organizations are wrestling with the same control question as they bring agents into editorial work (European broadcasters).

Seshadri describes a seed dataset expanded by models, a maker-checker setup where one model answers and another judges, and domain experts like sports-magazine editors helping define what serious fans ask. He also names the honest problem: who tests the judge? Models grading models can share blind spots, and at this scale small errors become large ones. It is the same principle behind calibrated decision models that route uncertain cases to humans (Gemini Decisions API).

Accuracy, latency and cost form a constant three-way tradeoff. With more than 60% of users choosing voice, answers must be fast and cheap as well as right. His regret, not investing earlier in monitoring model quality in production, is one other teams should take seriously.

Seshadri pushes back on vanity metrics, saying lines of code generated don’t prove value. The flip side is the risk: give engineers a powerful tool without the judgment to validate production systems, and risk multiplies along with output. That is the honest version of the AI productivity story, and it applies well beyond streaming. Similar unglamorous retrieval wins show up in industry (Kuken Industries). More in our OpenAI coverage and enterprise coverage.