Apple Park — "Meet Your New Best Friend"

SEA Hacks prototype: an AI buying experience for Apple Park Kids that reframes the purchase from "browse a catalog" to "meet the friend you're giving a child."

Flow: Landing → conversational gift-finder (4 questions) → 3 real matching dolls → pick one → personalized "Meet [Doll]" storybook page with streamed story + story video that ends on a cliffhanger → mock QR unlock → next-chapter teaser.

Run it

npm install
npm run dev

Open http://localhost:3000 and click Find your child's friend.

Environment

.env.local (gitignored):

INFERENCE_API_KEY=...   # Inference inference key (sk-inf-v1..., not sk-mgmt-v1...)
INFERENCE_BASE_URL=https://model.service-inference.ai/v1
INFERENCE_MODEL=deepseek-chat

⚠️ A management key (sk-mgmt-v1-...) can list usage/cost data but cannot run completions (401). Use an inference API key (sk-inf-v1-...) for the app. Until then, every route degrades gracefully to warm, hand-written fallback content — the demo is fully walkable either way.

Optional, for live Inference.ai Seedance video generation (off the demo critical path):

INFERENCE_VIDEO_MODEL=dreamina-seedance-2-0-260128
INFERENCE_VIDEO_DURATION=4
INFERENCE_VIDEO_RESOLUTION=480p
INFERENCE_VIDEO_RATIO=16:9
INFERENCE_VIDEO_AUDIO=false
NEXT_PUBLIC_SITE_URL=https://your-deployed-app.example.com
# Optional: reuse an existing asset group instead of creating one per request.
INFERENCE_ASSET_GROUP_ID=group-...

For a reference doll image, Inference.ai needs a public image URL. Local /dolls/*.png paths only work after deployment, when NEXT_PUBLIC_SITE_URL points at a reachable host. Without that, the live route still works as a text-to-video prompt based on the generated story.

If /video/generate returns 403 Model ... is not available to your account, the request format is fine but the model is not enabled for that inference key. Ask Inference.ai to enable dreamina-seedance-2-0-260128, or set INFERENCE_VIDEO_MODEL to a model your account can use.

How it's put together

• Next.js 16 (App Router) + TypeScript + Tailwind 4, Vercel AI SDK → Inference.ai (deepseek-chat).
• data/dolls.json — the 8 available "Park Friends" + Chloe, built from the real Shopify catalog (names, prices, personalities, product photos in public/dolls/). Ella is sold out and excluded.
• API routes (server-side only; the key never reaches the client):
  • POST /api/chat — streams each warm gift-finder question.
  • POST /api/recommend — generateObject (zod) → exactly 3 real dolls with reasons; keyword-match fallback.
  • POST /api/story — streams the personalized chapter-one story; template fallback.
  • POST /api/generate-video — live Inference.ai Seedance submit→poll→download proof.
  • GET /api/generate-video/status?taskId=...&dollId=... — poll a long-running video task and save the finished clip.
• public/videos/{id}.mp4 / {id}-full.mp4 — cached story clips (Ken Burns over real product photos, cliffhanger end cards). Regenerate with node scripts/gen_videos.mjs.

Demo script (fixed persona)

1. Landing → Find your child's friend.
2. Answer: Emma → 5 → birthday → shy and gentle, loves bedtime stories and tea parties.
3. Pick Wren → story streams starring Emma & Wren → play the video (ends "To be continued…").
4. Bring Wren home → mock confirm + outfit upsell (free shipping at $75).
5. Scan/click the QR → /unlock/wren → chapter-two clip → "Meet the next friend."