JioStar’s Chief Architect on AI Content Discovery at the 500-Million-User Scale | OpenAI
November 11, 2026
In conversation with Vijay Seshadri, Chief Architect at JioStar.

Company size: Enterprise
Region: APAC & Oceania
Industry: Media, Entertainment & Sports
Products: API, Codex
Our Executive Function series shares insights from leaders driving transformation with AI.
JioStar operates entertainment and live sports platforms serving over 500 million consumers, holding a record for supporting up to 72.5 million concurrent viewers during live sports broadcasts. At this scale, AI must be fast, reliable, and consistently production-ready.
Using the OpenAI API, JioStar built a conversational video and cricket stats discovery assistant directly into JioHotstar. Viewers can speak in natural language to find movies and shows, historical highlights, live match moments, or cricket statistics, with the system returning results through a visual, conversational experience.
At the OpenAI Executive Summit in Singapore, Sidharth Sharma, Head of Go-To-Market for ASEAN at OpenAI, sat down with Vijay Seshadri, Chief Architect at JioStar, to discuss how his team built an experience supporting over 15 languages in roughly eight weeks—and what leaders can learn from taking AI from demo to production.
What convinced you that AI would fundamentally change how people discover content?
Content discovery has barely evolved over the past 25 years. Recommendation systems were the last major shift, and most developments since have just been iterations on that same idea.
However, choosing what to watch is deeply personal. Someone might say, “I’m feeling sad; show me something that will cheer me up.” Even a request that simple carries a lot of nuance. What kind of sad? What usually cheers them up? Are they looking for a comedy, something heartwarming, or a complete distraction?
Traditional recommendation systems struggle to understand the needs of that exact moment. Conversation gives viewers a way to explain their state of mind. We call this a “voice-visual interface.”
“Content discovery is inherently visual, but combining those visual options with spoken preferences and conversation creates a powerful new paradigm.”
—Vijay Seshadri, Chief Architect, JioStar
How did you turn this idea into a product just in time for a major cricket tournament?
We started conceptualizing this in October and November, with a major cricket tournament coming up in February. We only had about two months, and we were dealing with two very different audiences: entertainment viewers and cricket fans.
An entertainment viewer might say, “It’s Friday night—surprise me,” or describe a movie plot without knowing the title. Sports viewers want much more specific action, like the fall of wickets, back-to-back sixes, or matches where a player scored a specific number of runs. Cricket experts might even want ball-by-ball statistics.
These are fundamentally different discovery problems. Working on a tight timeline, we focused on unifying entertainment content, sports highlights, and detailed cricket statistics into a single experience across multiple languages.
At JioStar’s scale, what trade-offs shaped your technical design?
Very early on, we framed the problem as a trade-off between accuracy, latency, and cost. Every engineering decision had to balance all three simultaneously.
Voice made this especially critical. Over 60% of our users interact via voice rather than text—people simply prefer to say what they want to watch out loud. At the same time, the system’s responses had to remain relevant, fast, and economically viable at massive concurrency.
“As cliché as it sounds, the answer is evals, evals, evals. If you’re building an application like this for the first time, even when you think you’re prepared, you will still likely underestimate the importance of evaluations.”
—Vijay Seshadri, Chief Architect, JioStar
What does an effective evaluation system look like in production?
I think about it in two parts: the evaluation framework and the validation layer. Before going to production, your evaluation framework should drive your iterations. Once in production, feedback loops should turn customer behavior and failure cases into new eval cases.
This flywheel becomes self-reinforcing, though getting it right is extremely difficult.
Domain experts are essential. For sports statistics evals, we partnered with professionals like sports magazine editors who understand the exact questions die-hard fans are likely to ask. You can’t just start building with a prompt alone; you need to seed the system with deep human expertise.