PinLanding: Turn Billions of Products into Instant Shopping Collections with Multimodal AI

Pinterest introduces PinLanding, a production pipeline that uses multimodal AI to automatically generate shopping collections from billions of catalog items.

• •User search history, autocomplete, and browse paths are analyzed to identify high-demand product spaces with thin collection coverage
• •A vision-language model (VLM) generates normalized key-value attribute pairs per product, then a curation pipeline applies frequency filtering, embedding-based clustering, and LLM-as-judge to build a compact attribute vocabulary
• •A CLIP-style dual-encoder model replaces per-product VLM inference at scale, achieving 99.7% Recall@10 on Fashion200K, far exceeding prior methods in the 50% range
• •Ray streaming jobs handle batch inference across millions of pins on 8 NVIDIA A100 GPUs (~$500/run), while Apache Spark constructs feeds via attribute-based ANN matching
• •Collections are evaluated through public benchmarks, LLM-as-judge, and human raters comparing against search-log-derived baselines