ToneWise

Practice real Mandarin conversations with an AI that grades your tones in real time — and drills your actual mistakes with spaced repetition.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for ToneWise

Inspiration

The single biggest reason native speakers can't understand a Mandarin learner isn't grammar or vocabulary — it's tones. Say the right syllable with the wrong tone and 水饺 (shuǐjiǎo, "dumplings") becomes 睡觉 (shuìjiào, "sleep"). Yet almost every learning app drills flashcards and never corrects your tones in an actual conversation. ToneWise fixes exactly that gap.

What it does

You pick a real-life scenario (ordering food, travel, a work meeting…) and your HSK level, then have an actual back-and-forth with Lin Wei, an AI native speaker. After every reply, ToneWise:

  • Grades your tones and grammar (1–10) with the specific tone errors it heard
  • Shows pinyin + a more natural phrasing
  • Ends each session with a summary (strengths, top mistakes, vocab to review)
  • Builds a spaced-repetition review queue from your actual tone mistakes, so you drill the things you personally get wrong

Which AWS Database, and how I use it

Amazon DynamoDB, with a deliberate single-table design (table tonewise):

Entity PK SK
User profile (plan, free_used, streak) USER#<uid> PROFILE
Session USER#<uid> SESSION#<id>
Graded turn SESSION#<id> TURN#<n>
Review item USER#<uid> REVIEW#<word>
  • A Global Secondary Index (GSI1) powers the "words due for review now" query — the heart of the spaced-repetition feature.
  • Every tone error from grading auto-enqueues into the review queue: a real write-then-GSI-read access pattern, not just a stored flag.
  • Pay-per-request billing means it scales to zero when idle and scales up automatically under load.

How I built it

  • Frontend + API: Next.js 16 (App Router) on Vercel, scaffolded with v0
  • Data: Amazon DynamoDB via AWS SDK v3 (@aws-sdk/lib-dynamodb)
  • AI: Google Vertex AI — Gemini 2.5 Flash for the opening line, per-turn tone+grammar grading, and the session summary (structured JSON). Authentication is keyless: a service-account JWT signed locally with Node crypto, exchanged for a short-lived OAuth token — no static API key in the request path.
  • Three API route handlers — /api/session, /api/chat, /api/summary — orchestrate DynamoDB + Gemini

Monetization (Track 1: B2C)

Free for 3 sessions, then a $5/month subscription for unlimited practice. Free-quota and subscription state live in the user's DynamoDB profile item, enforced server-side.

Accomplishments I'm proud of

  • A genuinely useful tone-feedback loop that turns each conversation into a personalized study plan.
  • A single-table DynamoDB model with a GSI that does real work (the review queue), not just CRUD.
  • A keyless path to the model provider, on a true zero-ops stack (Vercel + DynamoDB) that scales from one user to many without re-architecting.

What I learned

DynamoDB single-table modeling and when a GSI is the right tool, the v0 → Vercel workflow for shipping a full-stack app fast, and how to make an LLM a dependable, structured backend component.

What's next

Voice (speak your reply, hear Lin Wei via STT/TTS), leaderboards & streaks at scale, and more HSK content and additional languages.

AWS Database used: Amazon DynamoDB (single-table design + GSI). Live: https://tonewise-topaz.vercel.app

Built With

Share this project:

Updates

Submission history