Employment · 02
Stealth Startup
Co-founding a personalized skincare website with working product recommendations, client plans and routine logs. AI features included a RAG-powered chatbot and image analysis. Near MVP by August 2025, with classification still in development.
Explore a schematic mobile prototype and clinic workspace. The detection view uses illustrative scores, not patient data or measured model results. All portfolio information is available in the page text.
Mobile routine tracking and a clinic review workspace · illustrative prototype
The work
Personalized skincare, from product data to recommendations
Overview
From May to August 2025, I co-founded a skincare website for personalized product recommendations, client plans, and routine logs.
The business combined recommendations for existing skincare products with plans for our own line developed through an original design manufacturer (ODM). The website, recommendation platform, and product catalog were parts of that broader idea, with an iOS app as an additional client.
My responsibilities
I pitched to investors, led frontend development for the website, and built the product recommendation platform, alongside work on the iOS app and AI features.
Product data and recommendations
I collected skincare product information to support the recommendation platform and affiliate marketing. The goal was to connect what we knew about a client’s skin and needs with relevant products, including our planned skincare line.
AI-assisted routines and chatbot
I built AI-assisted routine tracking and a chatbot with a retrieval-augmented generation (RAG) response pipeline.
Image analysis and facial data
I worked on LLM-based image analysis, YOLOv8 acne and facial-feature detection, and image classifiers. I also worked on facial-data collection with privacy and applicable legal requirements in mind.
The training-data challenge
As an unfunded, pre-seed startup, our biggest challenge was building a reliable training dataset with limited resources. We researched image sources across the web, looking for licenses that permitted our intended use and examples supported by clinical assessments.
Finding images was only part of the problem. We also needed reliable labels for classification. A doctor’s assessment of a single photograph could be provisional: additional images of the same person might be needed to establish a more confident diagnosis.
The challenge was finding both usable images and enough clinical evidence to support their labels. That uncertainty made training-data quality a central constraint on our classification work.
Progress by August 2025
We were close to a minimum viable product. The recommendation platform worked, while the AI classification still needed further development.