JioTV+ Xtras
Drove 18% watch time increase by launching a contextual AI overlay akin to Amazon X-Ray — enriching passive viewing with real-time generative information.
The prototype runs on Streamlit's free tier and a scheduled check keeps it awake. If you still see a sleep screen, click the button and it is ready in about a minute.
The Challenge
Streaming viewers passively consume content without engaging deeper. There was an opportunity to transform passive viewing into active discovery by providing contextual information about what's on screen — similar to Amazon's X-Ray feature but powered by generative AI for Jio's massive content catalogue.
The Solution
- Designed JioTV+ Xtras as a contextual AI overlay that enriches the viewing experience with real-time generative information
- Built an experience akin to Amazon X-Ray but tailored for Indian content and Jio's platform
- Unlocked an estimated ₹110 Cr+ annual value opportunity by increasing viewer engagement and time-on-platform
- Integrated with existing JioTV+ infrastructure for seamless overlay rendering
- Built a working prototype demonstrating the core AI detection pipeline — classifies video segments into songs, dialogue, action, and credits using audio analysis (librosa) and vision language models
What the Prototype Produces
A demo run on a 63-second clip stitched from five film songs, cut into 10-second windows with 2 seconds of overlap and scored by audio features and a vision-language model together.
- 0:00–0:18 Musical 69% confident · sent to a human annotator Vision model: musical (67%); audio: musical (55%).
- 0:16–0:42 Dialogue 60% confident · sent to a human annotator Vision model: dialogue (67%); audio disagreed: musical (43%).
- 0:40–1:03 Musical 67% confident · sent to a human annotator Vision inconclusive; audio: musical (46%), 162 BPM, harmonic ratio 0.70.
Every segment scored under 70%, so instead of auto-tagging, the prototype routed each one to a human annotator, and it called the middle stretch dialogue while the audio said music. That is the case confidence routing is for: auto-tag what the models are sure of, send the uncertain stretches to a person, and keep a wrong label off the screen.
Impact
Key Decisions
- Contextual over intrusive: Designed the overlay to enhance rather than interrupt the viewing experience, triggered by user intent signals rather than auto-populating.
- Value-first framing: Modeled the ₹110 Cr+ annual value opportunity upfront to secure leadership buy-in before engineering investment.