The first concept
Skincare shoppers may know what concerns them without knowing which ingredients or product types to explore. My first concept asked users to upload a photo, then used image classification to identify skin characteristics and guide the conversation.
I assembled image datasets and trained and troubleshot models in TensorFlow and Keras to see whether the approach could produce dependable results.
Why I stopped pursuing it
The available data was too limited and inconsistent, and the model did not generalize well enough. Because the output could influence skincare choices, I did not think image classification was reliable enough to remain the core experience.
The pivot
01
Start with what users can state
The revised flow asks users to describe their concerns, goals, and preferences.
02
Retrieve relevant ingredients
I explored embeddings and vector search, then connected stated concerns with information from an ingredient knowledge base.
03
Show the reasoning
Responses explain why an ingredient may be relevant, giving users a basis for evaluating the guidance.
What exists now
BeautyGPT is a live conversational prototype supported by a PRD and user flow. It uses stated concerns and ingredient retrieval to answer skincare questions while keeping the basis for its guidance visible.
After the pivot, I could test the conversation, retrieval, and response quality directly. The prototype below shows the current experience.