Mukuntha Narayanan

Mukuntha Narayanan is a machine learning engineer whose work on Pinterest Search has made large-language-model relevance judgments practical for multilingual visual discovery. His research spans speech recognition, natural-language processing, computer vision, and the production systems needed to serve sophisticated models efficiently.

Publishing as Mukuntha Narayanan Sundararaman, he co-authored early research on crowd detection and counting and led a 2020 study of aspect-level sentiment transfer. At Observe.AI, he developed PhonemeBERT, a speech-aware language model that combines phonemes with automatic transcripts to reduce sensitivity to recognition errors. He also investigated how language models represent conversational structure and transcription mistakes. Narayanan earned a master’s degree in machine learning from Carnegie Mellon University, where he helped teach Machine Learning for Structured Data in 2022.

His subsequent Pinterest search research combines image-generated captions, Pin descriptions, search behavior, and user-curated context to judge whether visual content matches a query. At the 2025 AI Engineer World’s Fair, he and Han Wang explained their search-relevance system; Narayanan focused on making its strongest models economical enough for production.

• Teacher-student knowledge distillation: A multilingual teacher trained on human relevance judgments labels examples sampled from daily search traffic, transferring its capabilities into a smaller model while extending coverage across fresh content, markets, and languages.
• Production-scale bi-encoder search: Pins and queries are encoded separately: Pin embeddings are computed offline and refreshed when their inputs change, while short query embeddings are generated online and cached. This architecture reduces inference costs while retaining useful relevance signals.
• Multilingual relevance transfer: Training on global search activity helps propagate judgments initially concentrated in English-language annotations to other languages and regions.
• Reusable semantic representations: Relevance-trained embeddings can represent Pins, queries, and boards beyond search reranking, supporting adjacent discovery surfaces such as related content and home-feed recommendations.