Inspiration
People wonder about their skin right before a purchase, not in the abstract. Today that moment means a quiz, a guess, a serum that breaks you out, a lipstick that looked better on the model, and paying full price at the first store. Mira turns that moment into one text message: a measured skin read, products chosen for the measured problem, the shade already on your face, and the cheapest verified seller.
What it does
Mira is a shopping assistant you text like a friend. Send one selfie and Mira reads your skin with YouCam AI Skin Analysis, matches products to your three weakest scores, checks three stores for the best price, and then tries the best lip shade on your own face with YouCam Makeup VTO before you ask. Ask for a full look and Mira dresses your full body photo with YouCam AI Clothes. Everything happens inside an iMessage style thread that runs on iPhone Safari (add it to the home screen for a full screen app) and on desktop inside a phone frame.
Built for the YouCam API Skin AI and eCommerce VTO Hackathon (both tracks: Skin AI and eCommerce VTO, working as one flow).
How we built it
- Frontend: React and Vite, a pixel faithful iMessage thread (bubbles with tails, typing indicator, photo sheet, rich cards), responsive from a full screen iPhone web app to a phone frame on desktop.
- Agent: a client side planner (
useMira) turns each message or photo into a chain of tool calls and streams a live step trace into the thread, so the user watches the agent work: read photo, Skin AI, match, price check, Makeup VTO. - Backend: a dependency free Node server that keeps the YouCam key secret and implements the v2 flow for all three features: register file, PUT bytes to the presigned URL, create task, poll until success, and normalise the result shapes.
- YouCam APIs used: AI Skin Analysis (
skin-analysis, 8 SD concerns, JSON output with mask overlays), Makeup VTO (makeup-vto,lip_coloreffects with matte, satin and gloss textures), AI Clothes (cloth-v4with fallback tocloth).
Challenges
Each YouCam feature has its own input window (minimum short side for skin, maximum long side for makeup, full body pose for clothes), so the client re-encodes every photo to fit the feature it is about to call, and the server maps YouCam error codes (face too small, lighting too dark, pose) to friendly, actionable replies instead of a generic failure.
The most common rejection is a selfie taken at arm's length (YouCam needs the face to cover over 60% of the image width), so the agent heals itself: it zooms toward the face and retries once, and the step trace says so, before it ever asks the user for another photo.
What we learned
The best moment to use Skin AI is the moment right before a purchase, and the best way to sell a try-on is not to ask for permission: when the agent already has your face and your scores, showing the result unprompted turns a feature into a conversation.
What's next for Mira
A real catalog and price feed, YouCam Skin Tone analysis to pick foundation and undertone aware shades, weekly check in scans that track whether a product is working, and a native iMessage relay so the same agent answers texts sent to a phone number.
Built With
- ai-clothes)
- makeup-vto
- node.js
- playwright
- react
- typescript
- vite
- youcam-api-(ai-skin-analysis