Model Architecture

Core to construction is a robust infrastructure built to train across the world's commerce data and contextual signals.

Training Data

• ~15mm curated product images spanning ~5mm products from ~10k retailers
• Non-generic web data rather than commodity crawl coverage
• Taste-structured corpus, continuously refreshed via proprietary crawling infrastructure
• Weekly domain visits with catalog, style, and assortment change assessment
• Simultaneous cultural crawl of restaurants, menus, reviews, trend sources, and adjacent signals
• Cross-domain ingestion so the model learns taste across commerce, dining, and lifestyle