Turn your Chinese thoughts into natural English — then lock them in with spaced repetition.
Record · Input Chinese and generate natural English |
Save Flow · Review the generated text before saving |
Decks · AI auto-organizes cards into the right deck |
Review · Practice, rate, and reinforce with FSRS |
1. The complete loop no mainstream app has Type a Chinese thought → AI rewrites it as natural English → saved as a flashcard → reviewed via FSRS spaced repetition. Most vocabulary apps stop at word lists. SayRight closes the full cycle from expression to retention.
2. Group Agent: AI auto-organizes your cards When you save a card, an AI agent reads your existing decks and decides where it belongs — creating a new deck if needed. No manual tagging. Cards find their own home.
3. FSRS + AI review scoring During review, optionally type your own English attempt and get an AI score (0–100) mapped to Anki's four grades (Again / Hard / Good / Easy). The FSRS algorithm schedules the next review based on your actual retention — not fixed intervals.
| Layer | Technology |
|---|---|
| Frontend | React 19, TypeScript, Tailwind CSS v4, Vite |
| State & Data | TanStack Query, React Hook Form, Zod |
| Backend | Hono on Cloudflare Workers |
| Database | Cloudflare D1 (SQLite) via Drizzle ORM |
| Auth | better-auth (session cookie) |
| AI | Claude / GPT-4o (OpenAI-compatible) |
- Record — input Chinese (up to 200 chars), generate natural English, edit before saving
- Auto-deck — Group Agent assigns each card to the best existing deck, or creates one
- Review — FSRS scheduling with 4-grade self-rating; optional AI scoring with feedback
- Deck management — view cards by deck, move cards between decks, delete empty decks
- Home dashboard — today's review count, recent decks, streak stats
- Responsive — bottom nav on mobile, top nav on desktop (Warm Orange theme)
- Session auth — cookie-based auth via better-auth; multi-tenant data isolation
┌─────────────────────────────────────────────┐
│ React 19 SPA (Vite) │
│ TanStack Query · React Hook Form · Zod │
└──────────────────────┬──────────────────────┘
│ HTTP / JSON
┌──────────────────────▼──────────────────────┐
│ Hono · Cloudflare Workers │
│ better-auth · Drizzle ORM │
│ Group Agent (AI orchestration) │
└──────────────────────┬──────────────────────┘
│
┌────────────▼────────────┐
│ Cloudflare D1 (SQLite) │
└─────────────────────────┘
- Node.js 20+, pnpm
- Cloudflare account (for D1 and Workers)
- Wrangler CLI:
pnpm install -g wrangler
cd frontend
pnpm install
pnpm dev # http://localhost:5173cd backend-hono
pnpm install
# Create D1 database (first time only)
wrangler d1 create say-right
# Paste the database_id into wrangler.toml
# Run migrations
wrangler d1 migrations apply say-right --local
# Start local dev server
pnpm dev # http://localhost:8787Copy .dev.vars.example to .dev.vars in backend-hono/ and fill in:
BETTER_AUTH_SECRET=your_secret
LLM_MODE=deterministic
# deterministic only uses the built-in stub; switch to provider for real translation
# If you switch to provider mode, also fill these in:
LLM_API_KEY=your_key
LLM_BASE_URL=https://api.openai.com/v1 # or any OpenAI-compatible endpoint
- Real LLM integration for English generation — API contract defined, provider connection pending
- Group Agent with live AI — stub complete, wiring to provider pending
- AI review scoring with live model — stub complete, wiring pending
- Deploy to Cloudflare Workers + D1 (production)
- Voice input: speak Chinese → auto transcribe → generate English (v2)
- Pronunciation practice: read English aloud → AI scores accuracy (v2)
MIT
Live Demo: https://say-right.kaili.dev