Meesho reposted this
Before you scroll, Meesho’s systems have already processed 2 trillion signals to decide what to show you. In an episode of Decoding AI, Anagh Prasad, Investor at Accel, sat down with Debdoot Mukherjee, Chief Data Scientist & Head of AI at Meesho, to unpack what building AI at true Bharat scale actually looks like. Two ideas stood out for anyone building consumer AI in India. 𝐅𝐢𝐫𝐬𝐭: 𝐬𝐞𝐚𝐫𝐜𝐡 A huge chunk of Meesho’s searches don’t look like clean Google-style queries. Someone looking for a fan might type “hawa denewala machine.” Traditional search systems struggle with this because the exact keywords don’t match. LLMs work better because they can understand intent and context, even when the phrasing is unusual. 𝐒𝐞𝐜𝐨𝐧𝐝: 𝐯𝐨𝐢𝐜𝐞. Meesho is already handling 60,000+ voice calls a day, but the interesting part is why voice works so well for Bharat users. Typing vernacular languages on a Latin keyboard is hard, spellings are inconsistent, and keyword search assumes users already know how to “query” a system. The real challenge isn’t just better models, but reducing friction between user intent and the system, whether through LLMs that understand messy natural-language queries or interfaces like voice that remove the need for typed keywords entirely. The companies that win in India will build around how Bharat naturally communicates, not how users are assumed to behave. The full episode is linked below.