India's JioHotstar built voice search for shows and cricket in 8 weeks

The short version

• JioStar, which runs the JioHotstar streaming service in India, built an assistant that lets viewers ask out loud for a show or a cricket statistic.
• A small team built it for more than 15 languages in about 8 weeks, in time for a major cricket tournament.
• Its chief architect says the hard part was testing the AI's answers, and that most companies underestimate how much testing they need.

What JioStar built

JioStar's streaming and live sports services reach more than 500 million people. At its peak, it has served 72.5 million people watching the same live match.

Using OpenAI's models, the company added a conversational search assistant to JioHotstar. Viewers can describe what they want in their own words. The assistant replies with options on screen, so the experience mixes talk and pictures.

Vijay Seshadri, JioStar's chief architect, described the project in an interview OpenAI published on 11 November. He said content discovery had changed very little in 25 years. Recommendations were the last big shift.

His point was that choosing what to watch is personal. Someone might say they feel sad and want cheering up. A recommendation engine struggles with that. A conversation lets the viewer explain.

The team had two very different audiences to serve:

• Entertainment viewers, who might say "surprise me" or describe a film without knowing its title.
• Cricket fans, who want specific moments, like every six in a match, or detailed statistics about every ball.

What is actually new

The speed stands out. The team worked through the idea in October and November. It had roughly two months to build before a major cricket tournament in February.

More than 60% of JioStar's users speak to the app rather than type, and that shaped the design. That made speed and cost as important as accuracy. Every design choice had to balance those three.

Seshadri's main lesson was about testing. He said the answer is evals, the sets of test questions used to grade an AI's answers, and that teams underestimate them even when they think they are ready.

For the cricket statistics, JioStar brought in sports magazine editors to write the test questions real fans would ask. The team then used AI to create more variations from that starting set.

They also used two models with opposite jobs. One worked out how to answer a question. The other checked the answer, like a strict editor. That raised a new problem: someone also had to test the checker.

Scope came first. Because viewers can ask anything, the team began by deciding what the assistant should refuse to answer.

What it means for companies in Indonesia

Indonesia also has many languages. Any company with a large catalog faces the same discovery problem JioStar did. Streaming apps and online marketplaces are obvious examples.

The useful lessons are about process. A small team with a clear goal moved fast. Expert test writers, here sports editors, were essential.

Seshadri also warned leaders against counting activity instead of outcomes. Lines of code written by AI, he said, do not show whether you created value. He described AI as a chance to simplify how a business works, and called it a wasted opportunity if companies don't take it.

Asked what he would change, he said he would have invested earlier in watching how well the AI performs once it is live, not just whether the servers are up.

What to do next

• If customers struggle to find things in your app, test whether they would rather describe what they want out loud.
• Bring in real experts, from merchandisers to customer service leads, to write the test questions before you build.
• Decide what the assistant must refuse to answer before you decide what it should do.
• Track the quality of live answers every week, and turn each failure into a new test question.
• Judge the project by a business result like more viewing or more sales. How much AI was used is not the measure.