TwinAI
Remix any viral video in seconds. TwinAI turns any proven reference (Reel, TikTok, Short, YouTube) into a personalized, shootable blueprint — hook, script, shot list, captions, edit plan, and a publish schedule — in the creator's own voice. Reference-based creation, not a clipper. We copy structure, never content.
This is a fresh, real rebuild: a Vite + React + TypeScript frontend, a Supabase backend (Postgres + Auth + RLS), and Gemini-powered edge functions that do the actual AI generation server-side.
Rebuilt after the original source was lost. The earlier deploy was a frontend-only mock with no real backend/AI; this version replaces the faked generation with real Gemini calls, real auth, and real persistence.
Stack
| Layer | Tech |
|---|
| Frontend | Vite, React 18, TypeScript, Tailwind, React Router |
| Auth + DB | Supabase (Postgres, Row Level Security, Auth) |
| AI | Google Gemini (gemini-3.1-pro-preview) via Supabase Edge Functions — provider isolated server-side, swappable |
System shape: see ARCHITECTURE.md — the authoritative
end-to-end design (Vercel SPA → Supabase data plane → Edge Functions → jobs queue
→ VPS worker → satellites). For deploy steps see DEPLOY.md.
What's real vs. roadmap
Real now
- Email/password auth (Supabase) + per-user profiles
- 90-second creator-DNA onboarding stored in Postgres
- Real AI blueprint generation — the edge function calls Gemini with the creator's DNA + the reference and returns a structured, shootable plan (no templates, no mocks)
- Atomic credit ledger (spend on generate, auto-refund on failure)
- Generation history, all under RLS
Built since this list was first written
- Pulling/transcribing the actual reference video — the ingest pipeline + VPS worker fetch and transcribe real TikTok/YouTube/IG references (via Apify + Whisper).
- In-app AI editing — being rebuilt. The old auto-edit pipeline was removed (see
docs/ai-editor-rebuild-status.md); recording, upload and playback of finished videos still work, and a new one-click editor is on the way.
- Checkout rails — the billing edge function + webhook are built; paid plans are gated off behind the
PAYMENTS_LIVE flag until go-live.
Roadmap (honestly not built yet / partially)
- Direct auto-publish: YouTube one-click upload works via OAuth; TikTok/Instagram connect is wired but publishing is gated on each platform's content-API app review. Until then we produce a ready-to-paste caption + schedule.
- Flipping
PAYMENTS_LIVE / POSTING_LIVE on for production.
Local setup
npm install
cp apps/web/.env.example apps/web/.env.local # fill in VITE_SUPABASE_URL + VITE_SUPABASE_ANON_KEY
npm run web
Backend setup (Supabase)
- Create a Supabase project (free tier is fine).
- Apply the schema:
supabase link --project-ref <your-ref>
supabase db push # applies supabase/migrations/0001_init.sql
- Deploy the edge function and set its secret:
supabase functions deploy generate-blueprint
supabase secrets set GEMINI_API_KEY=...
# optional: supabase secrets set GEMINI_MODEL=gemini-3.1-pro-preview
- Put the project URL + anon key into
.env.local (frontend) — the function reads SUPABASE_* and GEMINI_API_KEY from its own secrets.
Deploy (frontend)
Vercel: import the repo, framework = Vite. Set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY as env vars. vercel.json already rewrites all routes to index.html for the SPA.
Project layout
apps/web/src/
lib/ shims re-exporting @twinai/shared (supabase client, API layer, brand, timeline)
context/ auth provider
components/ Nav, AppShell, v2 primitives, …
pages/ Landing, Auth, Onboarding, Dashboard, Result, History, Gallery, v2/ (Create → Building → Capture → Review)
packages/shared/ @twinai/shared — types, brand tokens, API layer, scene timeline
supabase/
migrations/ 0001 → latest (profiles, generations, jobs, RLS, RPCs, buckets, storage caps, oauth nonce, worker liveness, tour flag)
functions/ edge functions (generate-blueprint, start-dna, dna-poll, social, …)
worker/ VPS job worker (reference ingestion + DNA; ffmpeg + whisper)
Brand
Tokens live in tailwind.config.js and src/lib/brand.ts, mirroring the Brand & GTM cheat sheet: ink #07070A, cream #F6F1E9, amber #FFB347, coral #FF5B7B, teal #65E5D8; signature gradient amber→coral→teal at 135°; Geist/Inter type. Voice: creator-to-creator, direct and warm — never "guaranteed viral".
