Han Wang is a machine learning engineer specializing in visual-search relevance: determining whether an image actually answers someone’s query and measuring whether search results improve. Her work combines language models, visual signals, behavioral context, and human judgment to make search ranking and evaluation more reliable.
In 2024, Wang was first author of research on improving Pinterest search relevance and coauthored a Pinterest Engineering article describing that work. She and Mukuntha Narayanan subsequently presented their search research at AI Engineer World’s Fair 2025, where both were affiliated with Pinterest.
Wang’s work tackles several interconnected challenges:
- Visual-search relevance modeling: Cross-encoder language models assess query–Pin matches across five relevance levels, improving semantic judgment beyond conventional search embeddings.
- Vision-language captions and behavioral context: Generated image descriptions, Pin metadata, user-curated board titles, and previous search interactions collectively reveal both what an image contains and how people interpret it.
- Production-scale model distillation: Collaborative research transfers judgments from larger teacher models into smaller systems suitable for live search, extending relevance improvements across languages and markets.
- Automated relevance evaluation: As first author of 2025 research on search-quality assessment, Wang developed model-assisted evaluation grounded in human judgments, broader query sampling, and online experiments—an important safeguard when engagement alone obscures whether results satisfy user intent.
- Image-aware search experimentation: Her 2026 vision-language research evaluates visual results more directly and supports live A/B testing, bringing image understanding into the measurement process itself.
Her contributions span the full relevance cycle: representing visual content, ranking it against search intent, and testing whether changes genuinely improve the user experience.
