JioStar’s chief architect on AI discovery at 500M-user scale
JioStar’s conversational discovery launch highlights the work coding-agent teams need around evals, guardrails, domain experts, and production monitoring at scale.
JioStar built a conversational video and cricket-stat discovery assistant for JioHotstar in about **eight weeks** and across **more than 15 languages**. Its platforms serve over **500 million consumers** and have supported up to **72.5 million concurrent viewers**. For builders running coding agents, the interview stresses defining what the product should not answer, using domain experts to seed evals, and feeding production failures into new test cases. Codex sped up some implementation and eval work, while people remained responsible for validation. The material gives no measured model-efficacy or latency results for the assistant. The scale figures describe JioStar’s platforms, not necessarily traffic handled by the new feature.
No separate implication has been recorded yet.
JioStar built a conversational video and cricket-stat discovery assistant for JioHotstar in about **eight weeks** and across **more than 15 languages**. Its platforms serve over **500 million consumers** and have supported up to **72.5 million concurrent viewers**. For builders running coding agents, the interview stresses defining what the product should not answer, using domain experts to seed evals, and feeding production failures into new test cases. Codex sped up some implementation and eval work, while people remained responsible for validation. The material gives no measured model-efficacy or latency results for the assistant. The scale figures describe JioStar’s platforms, not necessarily traffic handled by the new feature.
This grounds agent-evaluation guidance in a multilingual product built for a very large platform, while clarifying that platform scale does not establish traffic or model quality for the new assistant. Its strongest contribution is an operating loop: define answer boundaries, seed tests with domain experts, and turn production failures into eval cases while people validate the results.