Role: Product Manager — Multilingual LLM & Indic NLP Products

Tenure: 2023 – 2025

Domain: AI/ML · Indic Language Technology · Open-Source Research Products · Government & Enterprise AI

Stage: Research-to-Production (0→1 applied product on top of open research)

At AI4Bharat I operated at the intersection of frontier NLP research and real-world product deployment — translating model capability into measurable accuracy, usability, and adoption outcomes for India's 22 scheduled languages.

Company & Context

AI4Bharat is a centre at IIT Madras, India — backed by Microsoft and NVIDIA — with the mission to build open-source, state-of-the-art AI for all 22 official Indian languages. During my tenure the org was building:

• IndicTrans2 — the first MT model facilitating translation across all 22 Indic languages and English (released May 2023)
• IndicVoices — 12,000 hours of speech data across 208 Indian districts and 22 languages (co-released with Sarvam AI)
• IndicASR — first ASR model supporting all 22 scheduled Indian languages
• IndicLLMSuite — 251B tokens + 74.8M instruction-response pairs across 22 languages for LLM training
• Airavata — Hindi instruction-tuned LLM
• Indic LLM Arena — evaluation benchmark for culturally-aware Indic language AI (supported by Google Cloud)

The core challenge: research-quality models existed but lacked the product layer — evaluation infrastructure, dataset quality pipelines, accuracy benchmarks tied to real use cases, and user-facing deployment.

The Unique Problem I Solved

Problem: Multilingual LLM accuracy was at 55% on real-world Indic benchmarks. Research models were being evaluated on English-centric leaderboards that masked catastrophic failure on Indian language tasks. There was no structured PM function to translate research outputs into product-ready systems with defensible accuracy metrics.

Why it was hard:

• No prior PM precedent for this kind of applied research product at an IIT lab