Biagiotti: An Integrated Intelligence Platform for Cosmetic Market Analytics Using Multi-Model ML

Abstract

The global luxury cosmetics market generates enormous volumes of fragmented product, ingredient, and sales data that remain largely inaccessible to individual retailers and dealers. Existing solutions address these domains in isolation, leaving a gap for end-to-end intelligence. We present Biagiotti, a full-stack market intelligence platform that unifies five machine learning pipelines—skin-type classification, sentiment analysis, ingredient safety detection, ingredient-based product similarity, and demand forecasting—within a single Flask-based web application backed by a 4,079-product catalogue. Our similarity engine extends the content-based filtering approach of Permana and Wibowo (2023), which applied TFIDF and cosine similarity to movie synopsis recommendation, and adapts it to the cosmetics domain using ingredient text as the similarity signal. Building on their baseline, we introduce bigram-aware TF-IDF, sublinear term-frequency scaling, and a sim ≥ 0.20 confidence filter that together substantially improve brand-level recommendation quality. Across all five pipelines, the system achieves sub-millisecond to lowmillisecond inference latency (mean 0.07 ms for similarity), 94.2% accuracy on harmful ingredient detection, Precision@1 of 90.0% with NDCG@10 of 91.5% for similarity, and an R² of 0.9186 on demand forecasting, making it suitable for real-time retail deployment. A dealer-facing web dashboard provides actionable insights across inventory, safety, sentiment, and demand in a unified interface.