Vijay Seshadri, Chief Architect at JioStar, said his team shipped a voice-led discovery assistant for JioHotstar in about eight weeks, as detailed by OpenAI News.
The assistant lets viewers speak naturally to find movies and shows, historical highlights, live-match moments or cricket statistics through a visual conversational result, built with the OpenAI API and Codex for more than 15 languages, per OpenAI News.
OpenAI News puts JioStar at more than 500 million consumers and as many as 72.5 million concurrent live-sports viewers, while TokenPost counts more than 500 million monthly users.
Sidharth Sharma, OpenAI's Head of Go-to-Market for ASEAN, questioned Seshadri at the Executive Summit in Singapore about shipping before a major cricket tournament in February after starting in October and November, per OpenAI News.
Inside AI described the same assistant as handling movies, live cricket moments and detailed statistics by voice or text, which matches the entertainment versus sports split Seshadri outlined.
Seshadri framed the design as a three-way tradeoff among accuracy, latency and cost because more than 60% of users prefer voice over text, per OpenAI News.
Seshadri summed up the lesson as evals, evals and evals.
He said effective work meant a pre-production evaluation framework plus a production validation layer fed by customer behavior, seeded by sports-magazine editors and expanded into a maker-checker pattern with an LLM as judge, per OpenAI News.
He said a small team of data scientists and engineers working with OpenAI forward-deployed engineers set guardrails first by deciding what the system should not answer, and used Codex to accelerate eval cases without removing judgment, per OpenAI News.
He said he would invest earlier in production monitoring for model efficacy because traditional P99 latency can't tell you if answers stay useful, per OpenAI News.
He warned executives not to mistake lines of code for outcomes and called AI a massive simplification hammer that demands skilled operators with validation and guardrails, per OpenAI News.
Who pays for voice inference when most viewers talk to the guide during a live final peak?
