StyloBuddy: A Hybrid Deep Learning and LLM Framework for Trend-Aware Outfit Recommendation

StyloBuddy is an outfit recommendation system that pairs deep visual compatibility scoring with large language model (LLM) reasoning and trend-aware catalogue enrichment. It has four components: (1) a GPT-4o Vision Pipeline for zero-shot clothing attribute extraction; (2) a Type-Aware Compatibility Network (TACN) trained on 15,000 synthetic outfit pairs from the Fashion Product Images dataset, which scores outfit compatibility through category-specific subspaces; (3) a Three-Mode Recommendation Engine that routes requests through three paths (online, auto-suggest, and offline fallback), with 97.3% prompt compliance; and (4) a prototyped Apify-based Trend Scraping Pipeline for periodic catalogue enrichment. The hybrid pipeline is deployed in production: for large item selections (6 or more items), TACN scores all upper-lower pairs and passes only the top-ranked combinations to GPT-4o, cutting LLM input by up to 75% while preserving recommendation quality. We report a comparative analysis (p < 0.01, Fleiss' kappa = 0.73), an ablation study, and a user study (N=45, mean satisfaction 4.12/5.0). This is a technical report and has not been peer reviewed.