Puse AI

Inspiration

What it does

How we built it

Challenges we ran in## Inspiration

Online beauty and fashion still force guesswork: shoppers buy skincare without knowing their actual skin metrics, and buy clothes without seeing the fit on their body. Returns and abandoned carts follow. YouCam’s Skin AI and Clothes Virtual Try-On APIs already solve each half of that problem — Puse chains them into one ecommerce path: diagnose → match → try on → bag → buy.

What it does

Skin AI (YouCam Skin Analysis)
Upload a selfie. Perfect Corp returns eight localized concern scores in one task (texture, pore, wrinkle, acne, redness, oiliness, moisture, radiance). Puse surfaces the weakest scores and the full diagnostic grid, then maps those concern types directly to skincare SKUs in the catalog (e.g. low pore → Pore-Refining Clay Mask). Matched products can be added to bag immediately.

eCommerce VTO (YouCam Clothes VTO v4)
Pick apparel, upload a full-body photo, and Clothes VTO renders the garment on you. From the result screen you add that look to the same bag (with the try-on image attached).

Unified shop loop
Product pages, bag, demo checkout, and a floating shop agent that chains Skin AI history to SKUs and Clothes VTO — one Puse-designed experience (not an open-source ecommerce template skin).

How we built it (Technological Implementation)

• Next.js 16 App Router — UI + backend in one app.
• Server-side Route Handlers (/api/upload, /api/skin-analysis, /api/try-on, poll routes) hold YOUCAM_API_KEY so the browser never sees Perfect Corp credentials.
• Shared custom polling hook (usePolledTask) — both Skin AI and Clothes VTO are async create-task → poll until success / error; one state machine, no racey re-uploads mid-poll.
• Multi-score → product chain — on Skin AI success we parse the full results.output[], sort by ui_score, take weakest concerns, and filter catalog products by concern field in real time (no second API round-trip for matching).
• Prisma 7 + SQLite — products, analysis results, try-ons, users, demo orders persist so History and checkout are real, not mocked.
• Cloth VTO uses documented ref_file_url + garment category (cloth-v4), not a fake template stub.

Live demo: hackathon.puse.ai

Challenges we ran into

YouCam tasks are asynchronous. We built a polling state machine that survives navigation and re-uploads cleanly. Cloth VTO needed the correct v4 contract (ref_file_url vs older template IDs). Mapping eight Skin AI types onto a seeded catalog without slowing the UI required keeping matching as a pure client-side filter over already-fetched products once scores arrive.

Accomplishments that we're proud of

A working dual-API ecommerce flow (Skin AI + Clothes VTO) with secure server-side auth, shared polling, concern→SKU matching, bag/checkout, and account-backed history — judged as a product, not a single “hello world” API call.

What we learned

Task-based AI APIs force different frontend architecture than REST CRUD. One selfie / one body photo becomes shared retail input across dermatology-grade scoring and generative try-on — that’s the real eCommerce unlock.

What's next for Puse

Larger multi-brand catalogs, skin-score trends over time, and production payment rails behind the same diagnose → try → buy UX.to

Accomplishments that we're proud of

What we learned

What's next for Puse AI

Built With

• nextjs
• prisma
• react
• sqlite
• tailwindcss
• typescript
• youcam-api
• youcam-clothes-vto
• youcam-skin-ai