plumchi - skincare is hard
skincare routine builder
skincare price comparison
skincare chat
plumchi perspective
progress tracker
plumchi logo 1024
skincare toutine
plumchi homepage
plumchi perspective
chat
product catalogue
progress checkin

Inspiration

I launched a small skincare brand as a side hustle called UXB (uxbskincare.com) in 2019. I learned how to formulate products, researched ingredients and skin conditions. I am deep in the world of beauty and aesthetics.

My main career was a product manager working at large broadcast media companies developing programme search and recommendation systems. I have a couple of patents around using machine learning to improve content discovery.

I have come to realise that finding the right skincare products us hard; you don't know what works for your skin, it's like a science experiment trying to weave all of the products into a routine you can follow and if you manage all that - you don't know if it's all working.

I have used my knowledge of skincare plus expertise in recommender algorithms to build the ultimate skincare assistant app.

What the plumchi app does

plumchi fixes 3 big problems:

1) You don't know what products will work for your skin 2) You don't know how to put products into a routine you can follow 3) You don't know if the products and routine is working for you.

After a short onboarding quiz we start recommending you products that suit your skin and fix your issues. Our routine builder then helps you put your products into a 7 day AM/PM routine that you can follow. Every day we'll prompt you to take a selfie and rate your skin so you can check progress.

We have thousands of products in the database - all with live prices.

If you need more skincare information you can use our chat to talk through your issues.

How I built it

Gotta be honest - leaned hard on Claude Code.

The backend has a data ingestion pipeline that takes product information and enriches it using LLMs and Tavily. All data is stored in Supabase (which also does authentication) and we have CloudFront fronted FastAPIs for all the CRUD operations.

The app is built using React Native and Expo. The paywalls are built and managed in RevenueCat (so easy to do).

The AI features are powered by various models through OpenRouter.

Challenges we ran into

I am finding it hard to get the app through Apple app review. Because the app verges on medical advice and uses AI features - data privacy has to be well thought out and declared.

Accomplishments that we're proud of

I put most thought into the chat. Skin conditions get people down and I wanted the chat to be empathetic yet grounded in reality.

I crafted a system prompt that guides the chatbot to be thoughtful and "understanding". It needed to know everything about skincare so I built an agent that uses Tavily to deep research on skin conditions and concerns and feeds it to the chat as retrieval augmented generations.

I went through 5 LLMs - testing using OpenRouter before I settled on Claude Sonnet 4.6 as the chat model. I tested/red teamed the chat using Ragas (retrieval-augmented generation assessment) and LLM as a Judge methodologies - and to be honest just talked to it a lot.

What we learned

80% of the time has been spent making sure the data is correct. Our classification and enrichment took a while before I was fully confident the data that the whole app relies on was solid.