Role: Product Manager — Data Strategy, Evaluation Systems & Applied AI Infrastructure

Tenure: 2024 – Jan 2026 · Domain: Indic NLP · Speech · LLM Pretraining Data · OCR · Government AI

Org: AI4Bharat, a center at IIT Madras backed by MeitY/Bhashini, Nilekani Philanthropies/EkStep, Microsoft, Google, CDAC and Yotta

I owned the data-to-model pipeline across five modalities (ASR, LLM, TTS, OCR, Transliteration) for India's national Indic-AI mission — deciding what data gets collected, what "good enough" means before a model can train on it, and how each decision cascades into benchmark-grade model quality.

Problem

96% of AI is built for English. India's 22 constitutionally recognized languages — including low-resource ones like Bodo, Dogri, Kashmiri and Maithili — had no production-grade AI support. The gap wasn't incremental; it was absolute:

Modality State before this work
ASR Existing datasets covered ≤5 languages, <100 hours each
LLM Indic text represented <1% of global pretraining corpora; tokenizers poorly optimized for Indic scripts
TTS No expressive, multilingual dataset existed for any Indian language — monotone, single-speaker only
OCR Indic document-layout parsing entirely unsolved; no labeled dataset across scripts/domains

MeitY's Bhashini mission mandated open-source, deployable AI for all scheduled languages — no vendor had delivered it. Sarvam AI and Microsoft Azure AI were entering the same space, validating the need but raising the bar on quality and breadth.

Approach & Framework

I ran a 4-phase, data-first methodology applied consistently across all five modalities:

  1. Decide — data strategy: which domains, languages, and sources to collect (field recordings vs. YouTube-scraped vs. studio), and what "good enough" means per modality (SNR thresholds for ASR, emotion-authenticity scoring for TTS, layout-class taxonomy for OCR, perplexity/toxicity filtering for LLM text).
  2. Build — collection-system architecture, not just protocol execution: integrated Microsoft Karya (crowdsourcing), Shoonya/Label Studio (annotation), and the Apache Spark-based Setu pipeline (web-text cleaning) into one pipeline per modality.
  3. Validate — inter-annotator agreement thresholds, rejection criteria for noisy/ambiguous samples, and benchmark evaluation gates (Vistaar, IndicSUPERB, IndicGLUE) before anything shipped.
  4. Evaluate — connected daily data decisions to downstream model quality via peer-reviewed benchmarks (ACL, INTERSPEECH, ICDAR).

Grounded in field discovery: recording operations across 400+ districts, 28+ states, direct exposure to the real end users (farmers, healthcare workers, government-service users who need voice-first, native-script interfaces — not English transliterations) and the full funder/partner ecosystem (MeitY, Nilekani Philanthropies, Microsoft, Google, CDAC, IIT Madras/IIIT Hyderabad).

Output — What Shipped, By Modality

Modality My Contribution Artifact
ASR Designed the audio collection architecture (field + YouTube + studio) and QC/validation pipeline (SNR thresholds, speaker-diversity requirements, inter-annotator agreement) IndicVoices — 23.7K hrs, 51K speakers, 400+ districts, 22 languages → IndicASR, first ASR model covering all 22 scheduled languages
LLM Designed inclusion criteria and filtering strategy for the Setu web-crawl subset (language ID, quality scoring, perplexity filtering); contributed to IndicAlign instruction curation and IndicGLUE evaluation Sangraha — 251B tokens, 22 languages → Airavata, Hindi instruction-tuned LLM; ACL 2024 Outstanding Paper Award
TTS Designed the collection architecture and emotion-validation pipeline (speaker criteria, recording protocols, inter-annotator agreement for subjective emotion labeling) Rasa — first multilingual expressive TTS dataset for an Indian language (6 Ekman emotions, 3 languages), accepted at INTERSPEECH 2024 → fed IndicF5 (11 languages, 1,417 hrs)
OCR Designed document-collection strategy and annotation validation pipeline (domain/source selection, 42 layout-class definitions, inter-annotator protocols across 12 document types) IndicDLP — 119,806 images, 12 domains, 11 languages + English; ICDAR 2025 Best Student Paper Runner-Up
Transliteration Designed source selection, quality thresholds and validation protocols for Roman-script Indic text Setu-Transliterate, open-sourced in IndicLLMSuite — unlocked romanized web data as usable LLM training signal across 14 languages

Outcome