At Epica Beauty we kept seeing the same question from users: "what face shape do I have?" People were measuring their faces with rulers and comparing themselves to celebrity photos. So we built a small tool that answers it from a single selfie: testmyface.com.
The stack is deliberately boring. The site is static HTML on Vercel. When a user uploads a photo, a single serverless function sends it to Claude's vision API with a prompt that asks for facial proportions (forehead width, cheekbone prominence, jawline shape, length-to-width ratio) and a classification into one of six shapes: oval, round, square, heart, diamond, or oblong. The response comes back as JSON with a confidence score and styling suggestions, and the frontend renders it. One page, one function, about $0.003 per analysis.
Photos are never written to disk. The image goes from the browser to the function to the API and the result comes back in a few seconds.
Things that surprised us: getting the model to commit to one shape instead of hedging between two took more prompt iteration than anything else, and a cheap small model classified faces as reliably as the larger ones once the prompt was strict enough.
It is free and has no signup. If you try it, I would love to hear whether the classification matches what you thought your face shape was.