Live app: https://you-mirror.vercel.app · Code: https://github.com/fozagtx/YouMirror

Inspiration

Every foundation looks right on the product page. Then it arrives two shades too light. Shoppers can't try a shade on their own skin from a screen, so they guess, send back what doesn't match, and stop trusting the page. No store can put a shade expert next to every shopper, so we built a mirror that answers one question right before you pay: should I buy this?

What it does

YouMirror turns a product page into a buying decision.

  1. Pick a real foundation or lipstick from 8,468 catalogue shades (464 product lines from Sephora and Ulta), or paste a garment link.
  2. Take one selfie. The app checks the photo locally first, so a blurry or too-small photo never spends YouCam units.
  3. YouCam Skin Tone Analysis measures your tone and undertone. YouCam Skin Analysis reads today's skin: acne, redness, oiliness, moisture, texture and radiance.
  4. YouCam tries that exact shade on you.
  5. The mirror answers Buy it, Skip it, or Try this instead, with reasons from your own scan. The two closest shades in the same line are tried on you too, each shown next to its real product photo, brand, price and store link.
  6. On a breakout day it raises coverage and adds a skincare step, so the advice changes with the day you're having.

How we built it

Next.js 16 on Vercel. One server route runs an agent loop and streams each step to the page: photo check, tone check, skin scan, try-on, compare, verdict. It calls four YouCam APIs: Skin Tone Analysis, Skin Analysis, Makeup VTO and Clothes VTO. MediaPipe finds the face in the browser and the server crops a portrait around it, because YouCam's scans need the face across about 60% of the frame. The catalogue comes from real Sephora and Ulta shade data, and a script checked every product photo link, giving 6,418 verified per-shade packshots. Built test-first with Vitest (57 tests).

Challenges we ran into

  • Wide webcam photos came back from YouCam as "face too small" or a false "face turned left". We fixed this with a face crop, plus a fallback to the full photo when the crop is too tight. Failed tasks cost no units.
  • YouCam slightly recolours the whole photo on every try-on, so a naive "unchanged" check flagged subtle nude shades. The honest unchanged-result check now runs on garments only: it retries once, then says so instead of showing a bad result.
  • Retailer catalogue images were mostly flat swatches or broken links, so we probed and verified real packshots.

Accomplishments that we're proud of

We ran our own skin-tone test: light to very deep skin, as shot, 45% darker, and under warm tungsten light. The reading moved by 1.4 to 4.2 ΔE, about the gap between two neighbouring shades. Very deep skin (L* 33) was read with no errors. Every rejection came from framing, never from skin depth.

What we learned

A try-on alone is a filter. Measured tone plus today's skin plus a try-on of the closer shades is a decision.

What's next for YouMirror

An embeddable "Ask the mirror" button for retailer product pages, more categories (concealer, blush), and garment fit using the same buy, skip or swap answer.

Built With

Share this project:

Updates

Submission history