Palta reposted this
CEO & Co-Founder at Lóvi | AI Innovator | Building the Future of Personalized Beauty Tech 🌐
Today, we rolled out updated skincare recommendations for all Lovi users! These updates leverage our new methodology for evaluating the personalized match between individual skin traits and cosmetic formulations 🚀 Finding cosmetics that truly work is a challenge for many skincare consumers. Before joining Lovi, I spent time exploring reviews on e-commerce websites and was struck by the sheer number of negative reviews, many of which could be attributed simply to the absence of reliable methods for determining whether a product meets an individual’s specific needs — whether it’s addressing particular skin concerns, or matching their skin type and sensitivity. At first glance, it might seem simple to find a product tailored to specific skin traits, given that the active ingredients beneficial for various skin conditions are well-known. However, analyzing individual ingredients is just the beginning — the real challenge lies in understanding the formulation as a whole. When assessing formulations, we consider often overlooked factors, including ingredient stability, compensatory effects for aggressive actives, the effectiveness of delivery systems, texture considerations, and more. Accounting for all these details is challenging, even for skincare professionals. To address this complexity, we developed a structured methodology that organizes all these factors into a clear framework. This methodology is now consistently used to determine matching scores, which are divided into four clear scales: - Safety - Functionality - Skin issue suitability - Skin type suitability For each scale, we developed a detailed evaluation guide. The methodology allows our experts to provide more consistent assessments and is a foundation for developing algorithms to replicate the same process on scale. Another challenging task was creating algorithms capable of scaling this methodology. I’ll probably cover the technical aspects in a separate blog post, but for now, here’s a brief overview: we used a combination of LLM and conventional ML methods, as well as some computational tricks. This approach achieved LLM-level generalization, ML-level alignment with expert opinions, and scalability to serve millions of users. In the meantime, check out the updated programs in the app and the cosmetic scanner, which now provides detailed Lovi Scores for recognized products. Try it out, and we’d love to hear your feedback! We plan to make our methodology public in the future, but for now, we’d be happy to share it with skincare professionals interested in validating and refining it. PS: The images show how the calculated fit score for the same product varies based on the user’s unique traits and how we incorporate these scores into personalized programs.