Meesho announced the integration of PRISM (Personalised Ranking & Intent Signal Module), an internal AI system handling product recommendations for its 264 million annual transacting users. While traditional e-commerce relies on keyword search bars, PRISM tracks real-time user behavior to predict intent. This discovery-based feed now generates over 75 per cent of the platform’s total orders.
The architecture uses over 100 AI ranking models trained on 400 trillion input signals. It executes 6 trillion daily inferences, which increases to 100 million inferences per second during peak traffic periods. To process these workloads, Meesho developed BharatMLStack, an in-house machine learning infrastructure designed to reduce data processing costs compared to standard cloud options.
PRISM also includes an LLM component called Trendpulse, which tracks regional purchase patterns to display inventory matching local demand. The system supports 10 languages, focusing on voice-led navigation and regional dialects. Debdoot Mukherjee, Chief Data Scientist at Meesho, stated that incoming internet users prefer browsing and voice commands over text search, requiring systems that interpret contextual behavioral data rather than explicit search queries.
The system currently manages product recommendations, trend analysis, and seller distribution metrics. It operates across a dataset generated by 263 million monthly active users, 17 billion daily product views, and over 2 billion combined user ratings and reviews.
Meesho posted NMV (Net Merchandise Value) of ₹11,371 crore in Q4 FY26, up 43 per cent y-o-y, with ₹717 million orders at the same growth rate. Losses narrowed 66 per cent as contribution margins rebounded to 4.0 per cent of NMV. Vaani, the voice shopping agent launched mid-quarter, crossed 1.5 million users within its first month and delivered a 22 per cent conversion lift.