MirrorRun

What it does

MirrorRun starts with the moment a retail team actually worries about: a shopper scans a QR code, opens a try-on path, and expects the experience to work. The operator sees that same session in a launch room with the media result, provider status, agent plan, fault controls, and a proof receipt tied to one session id.

When a dependency is missing or unhealthy, MirrorRun records the blocker instead of inventing a good-looking result. With sponsor keys installed, the same loop turns into live provider evidence: Perfect Corp produces the consumer-facing try-on output, Crusoe-hosted Nemotron writes the launch recommendation and risk checklist, TrueFoundry captures resilience behavior, and Lark turns the session into a replay workflow for QA.

How we built it

The app uses Next.js App Router with Cloudflare Workers/OpenNext so the UI, route handlers, storage adapters, and streaming AI calls share one deployment target. Development state is local JSON/media; the production plan uses Cloudflare D1 for sessions, R2 for media, and KV for lightweight provider status.

The visible product has four surfaces: the public landing page, the operator launch room, the phone-first shopper path, and the replay/detail views. Playwright covers the desktop and mobile flows, and the HackathonHunter visual QA scanner checks the actual browser for overflow, broken media, console errors, and small tap targets before screenshots or recording.

Challenges we ran into

The hard part was keeping the project honest while several sponsor platforms require account creation, email verification, or one-time API key reveal screens. MirrorRun treats those as product evidence, not something to hide: each provider panel can show either a real job/request/workflow id or a credential blocker with the official setup route.

Accomplishments we're proud of

The project is not a single demo page. It has a shopper route, operator route, replay route, architecture route, Cloudflare deploy plan, bilingual docs, a 5-slide deck, image assets, and browser-level visual QA. The build currently passes typecheck, production build, six Playwright tests, production smoke, and a real upload smoke that stores a 2.7 MB shopper image and attaches a Perfect Corp result URL.

What we learned

Retail AI demos need two kinds of proof at the same time: the shopper has to see a good result, and the launch team has to know what happened when a provider stalls. Keeping those two views tied to the same receipt makes the product more useful than either a try-on toy or an ops dashboard by itself.

What's next for MirrorRun

Sponsor credentials are installed locally and in the deployed Worker; provider panels now show configured status and attach either a real provider receipt or the exact provider error returned by the live endpoint.

Built With

• cloudflare-d1
• cloudflare-kv
• cloudflare-r2
• cloudflare-workers
• crusoe-managed-inference
• framer-motion
• lark-cli
• mantine
• next.js
• nvidia-nemotron
• opennext
• perfect-corp-youcam-api
• playwright
• radix-themes
• react
• tailwind-css
• truefoundry-ai-gateway
• typescript