At OpenAI’s Executive Summit in Singapore, Vijay Seshadri, Chief Architect at JioStar, described how the company added a conversational voice-and-visual discovery assistant to JioHotstar. JioStar’s entertainment and live-sports platforms serve more than 500 million consumers and have supported as many as 72.5 million concurrent live-sports viewers; at that scale, AI systems must be fast, reliable, and production-ready.
What they built
Using the OpenAI API, JioStar implemented a conversational assistant that lets viewers speak naturally to find films and shows, historical highlights, live-match moments, or cricket statistics. Results are delivered through a visual, conversational interface.
The feature was developed for more than 15 languages in roughly eight weeks. The concept work began in October and November because a major cricket tournament in February created a clear deadline.
Two audiences, one product, tight timeline
The project had to serve two distinct audiences: entertainment viewers and cricket fans. Entertainment users might say things like “It’s Friday evening—surprise me,” or describe a film without knowing the title. Sports viewers tend to ask for specific updates — fallen wickets, a sequence of sixes, or which match a player scored a particular number of runs. Cricket experts may request statistics down to every ball.
These are different discovery problems; the team’s goal was a single experience that handles general entertainment discovery, sports moments and detailed cricket statistics across multiple languages within a short time frame.
Engineering trade-offs: accuracy, latency and cost
Early on the team framed the challenge as a three-way tradeoff among accuracy, latency and cost. Every engineering choice had to balance those factors. Voice interaction made this especially important: more than 60% of JioStar users prefer voice over text, so responses must be relevant, fast and economically viable at scale.
Evaluation and validation in production
Seshadri split the approach into an evaluation framework and a validation layer. Before production, the evaluation framework should drive changes; in production, a feedback loop should convert customer behavior and failures into new evaluation cases. That flywheel can amplify improvements but is difficult to implement correctly.
Domain experts were essential. For sports-statistics evaluations, the team worked with people such as sports-magazine editors who understand what serious fans might ask. Seshadri emphasized that you cannot start with prompts alone; the system needs to be seeded with deep human expertise.
Team, tooling and patterns
The initial team was small — a handful of data scientists and engineers around a clear problem statement. OpenAI’s forward-deployed engineers collaborated with them to identify the critical items that had to be right.
Scope control came first: because a voice or text interface invites nearly any user question, the team decided early what the system should not answer and built guardrails accordingly.
For evaluations they began with a seed dataset and used models to expand it with more variation. They used a maker-checker pattern: one model helped determine how to answer a question while another model acted as a validator — an LLM as judge with a competing objective. Designing that validator raised another question: who tests the judge?
The team used Codex to accelerate work such as generating evaluation cases and updating the system, but Seshadri noted that faster building does not remove the need for judgment and validation.
Recommendations for leaders
Seshadri recommended investing earlier in production monitoring for model efficacy. Traditional infrastructure metrics like P99 latency remain important, but are not enough: teams must monitor model efficacy in production.
He highlighted three practical points: most companies lag behind every frontier-model announcement; leaders should focus on outcomes rather than visible activity (for example, generated lines of code do not equal value); and enterprises underestimate the human skill and agency required. Generic reskilling is not sufficient — leaders need people who can apply these tools with judgment. Give a skilled engineer a tool like Codex and impact can multiply; without judgment and validation the risk multiplies too.
"You have a rocket in your hands," Seshadri said — emphasizing that production systems need validation, guardrails and the right people holding the keys.
Summary
In less than three months JioStar shipped a multilingual, voice-driven conversational discovery experience into JioHotstar using the OpenAI API, combining entertainment discovery, match moments and detailed cricket statistics. Key lessons from the project include involving domain experts, establishing guardrails and an evaluation–validation loop, and investing in production monitoring and human judgment to operate AI at scale.
