Skyntri: An Integrated Web-Based Skin Health Assistant Using Multi-Label EfficientNetB0 Classification of Visible Skin Concerns

Skyntri is a web-based assistant that detects ten visible skin concerns (acne, blackheads, dark spots, dry skin, eye bags, normal skin, oily skin, pores, skin redness, wrinkles) from one facial photograph and links the result to practical skin-care support. It is an informational tool and does not diagnose disease. The image model is EfficientNetB0, pretrained on ImageNet and fine-tuned in two stages with TensorFlow/Keras as a multi-label classifier on the public Roboflow Skin-Problem-MultiLabel dataset of 4,828 facial images. Class imbalance was handled with a weighted focal loss and per-class decision thresholds tuned on the validation set. On an independent test set of 481 images, the model reached 71.52% subset accuracy, 81.17% micro F1, 66.73% macro F1 and 4.03% Hamming loss, outperforming a MobileNetV2 baseline on subset accuracy and every F1 measure. The model is served by a FastAPI backend to a React frontend, alongside a Tesseract OCR ingredient scanner matched against 167 curated ingredients, a safety-aware assistant, rule-based care recommendations, progress comparison and PDF reports; all 26 automated tests passed. The paper also specifies two planned extensions: skin disease screening and affordable product recommendation.