ASTER 2.0 was inspired by a simple but powerful idea: most people struggle to understand their skin and often buy skincare products through trial and error. This leads to wasted money, frustration, and sometimes worsening skin conditions. We wanted to build an AI assistant that could analyze a selfie, identify visible skin characteristics, and provide personalized skincare recommendations instantly.
Our goal was to combine computer vision and large language models to make skincare analysis accessible, understandable, and actionable. Instead of offering a medical diagnosis, ASTER 2.0 focuses on visible features such as skin type, acne, redness, pigmentation, pores, wrinkles, and dark circles. Based on these observations, the system generates a personalized beauty report and recommends real products already available on the market.
We built the project using a modern AI stack. The frontend was developed with Next.js, allowing users to upload facial images and receive interactive reports. The backend was built with FastAPI and handled image processing, model inference, and communication with the language model. For computer vision, we trained a deep learning model using multiple skincare datasets from Kaggle, including facial skin condition, acne, pigmentation, pores, wrinkles, and redness datasets. We used transfer learning with EfficientNet to classify skin characteristics. We then used Gemma to generate natural-language explanations and skincare routines tailored to each user.
One of the biggest challenges was data preparation. The available datasets had different structures, inconsistent labels, and varying image quality. We had to merge several datasets, clean invalid images, standardize labels, and create a unified training dataset. Another challenge was ensuring that the project remained ethically responsible. We designed the system as a beauty and wellness assistant rather than a medical diagnostic tool and included disclaimers advising users to consult dermatologists for serious concerns.
Throughout the project, we learned a great deal about dataset engineering, computer vision, transfer learning, prompt engineering, and product recommendation systems. We also learned how to connect vision models with large language models to create an end-to-end user experience that feels intelligent and highly personalized.
ASTER 2.0 demonstrates how AI can transform a simple selfie into practical, personalized skincare guidance. We believe this technology can help users make better decisions, reduce unnecessary product purchases, and better understand their skin. This project also has strong potential to evolve into a consumer beauty platform and a scalable startup.
Log in or sign up for Devpost to join the conversation.