AI-Powered Skin Analysis and Recommendation System

Abstract

This thesis presents an AI-powered system that produces personalised, safety-checked skincare guidance from a single facial photograph and a short questionnaire. The system is organised as a pipeline of seven cooperating agents built on a React and FastAPI stack, in which a multimodal large language model (GPT-4o) performs the visual perception and writes the final report, while every safety-bearing decision, the detection thresholds for skin concerns, the checking of conflicts between recommended products, and the referral of cases that require a professional, is handled by deterministic code. Recommendations are drawn from a real product catalogue, so that every suggestion can actually be obtained. The system is evaluated on a labelled test set spanning Fitzpatrick skin types III to VI, with particular attention to fairness across skin tones. The results show strong within-one-type classification accuracy alongside a systematic under-detection of concerns, reported openly rather than concealed.