JioStar built a 500M-user AI discovery assistant in eight weeks alongside OpenAI's forward deployed engineers

JioStar, the streaming company behind JioHotstar, built a conversational assistant that lets viewers find films, live cricket moments and detailed match statistics by voice or text. According to chief architect Vijay Seshadri, speaking with OpenAI's ASEAN go-to-market head at OpenAI's Executive Summit in Singapore, the system was developed in roughly eight weeks across more than 15 languages, ahead of a major cricket tournament in February.

The build started with a small group of JioStar data scientists and engineers working alongside OpenAI's forward-deployed engineers. Their first decision was scope: deciding what the assistant should not answer, so guardrails came before features. The team framed every choice as a tradeoff between accuracy, latency and cost, used a maker-checker setup where one model answers and a second model validates, and brought in sports-magazine editors to write realistic evaluation cases. Codex was used to speed up work like generating eval cases.

Seshadri said that in hindsight he would have invested earlier in monitoring how well the model performs in production, not just infrastructure metrics like latency.

For FDEs, this is a clear picture of what an embedded lab engagement looks like in practice: a fixed launch deadline, scoping and guardrails up front, evaluation design with domain experts, and production monitoring as the work that continues after launch.