As a personal care company, P&G is obsessed (in a good way) with understanding skin — what people care about, what actually changes skin health, and how well products deliver on those things. Most of the skin-related work I’ve done fits into a few recurring questions:
- What aspects of skin appearance actually matter to consumers?
- What can we realistically influence or improve through products?
- How do we measure product performance in ways that align with consumer perception and clinical science?
Across these problems, my role has been to build computer vision and machine learning tools that help answer them at scale.
Understanding What Matters to Consumers
When the answer isn’t already clear, we start by running perceptual studies with consumers. To support this, I’ve built ML and computer vision pipelines that generate synthetic images of skin with controlled variations in appearance. These let us probe which visual features — and which levels of severity — people are most sensitive to.
Working closely with statisticians and consumer researchers, we use tools like conjoint analysis to roughly rank skin appearance features based on consumer importance. This gives us a grounded, data-driven view of what actually matters, rather than relying on intuition alone.
Quantifying Skin Appearance
Once we have a coarse understanding of consumer priorities, I collaborate with clinical scientists and product designers to translate those priorities into measurable quantities. The goal is to define appearance metrics that:
- Matter to consumers
- Are biologically meaningful
- Map to mechanisms clinicians can actually intervene on
Together, we’ve developed several internal methods for quantifying the appearance of skin conditions that align well with known biology while remaining robust across confounders like skin tone. A lot of my work has lived at this boundary between perception, physics, and biology.
Measuring Product Impact
When a potential intervention or product is proposed, we run clinical studies to evaluate its effectiveness. During these studies, we track changes using our internally developed appearance metrics alongside established third-party baselines.
Once data is collected, analyses are validated by statisticians who weren’t involved in the study design, helping reduce bias and conflicts of interest. This process is repeated across products, treatments, and markets.
Scaling the Process
A major theme of my work has been making this entire pipeline faster, cheaper, and more scalable. For example, I’ve worked on methods that allow clinical-quality skin analysis using images captured at home — often just a selfie — rather than requiring in-clinic imaging.
We’ve also explored how treatment response varies across phenotypes, and shown that it’s possible to predict product response using an image of a subject’s face alone.
I will be providing more technical details on these projects soon.....