Mi Español Plus

Talk with an AI partner that corrects your Spanish in Chinese, then pay for the shirt in pesos with PayPal. Five minutes a day, for learners in Taiwan.

Comment

Inspiration

Self-taught Spanish learners in Taiwan can find flashcards, grammar notes and videos. What they cannot find is someone to talk to: someone who waits while they type a sentence, answers at their level, and explains the mistake in Chinese. A private tutor could do that, but a tutor is paid by the hour.

I am learning Spanish myself. Before this build I made Mi Español, a free site with a five-minute daily routine. All of its checking is rule-based. It knows whether a fill-in answer is right, but it cannot hold a conversation. For the hackathon I added the conversation, and one of the first real things a learner has to do in Spanish abroad: pay.

What was there before October 1

Mi Español was already live before October 1, with dialogue practice in 7 everyday scenes, spaced-repetition flashcards, listening and speaking practice, and progress tracking. It had no AI and no payments.

In early October, on the original site and not for this hackathon, I added the daily five-minute routine (October 6), seven more scenes for 14 in total (October 9), and fill-in items and A2 flashcards with example sentences from Tatoeba (October 6–11). Everything in the next section was built for the hackathon, from October 19, 2026.

What it does

  1. AI partner (Claude Haiku 5.5). A fourth mode in the conversation tab. Pick one of the 14 scenes and talk freely. The model plays the other person in simple Latin American (Mexican) Spanish and gives a Traditional Chinese translation of each reply. When the learner makes a real mistake, a card under their message shows the corrected sentence and a one-line reason in Chinese. Corrections go into the existing mistake book for later review.
  2. Mi Español Plus (PayPal Subscriptions). Free learners get 3 AI conversations a day, 20 turns each. Plus costs NT$60 a month and allows up to 200 turns a day. Learners subscribe and cancel on the progress page.
  3. Pay in Spanish (PayPal Orders, agentic). In three shop scenes (clothes shop, café, market), the AI shopkeeper takes the order in Spanish. When the customer is done, the model calls a create_bill tool. The server turns that call into a PayPal order from a fixed price list, and the page prints a ticket with a PayPal button. After payment, the shopkeeper thanks the customer in Spanish and the ticket gets a PAGADO stamp.

How I built it

  • The page is the existing single static file: HTML, CSS and plain JavaScript, no framework.
  • The back end is Vercel functions in Node. The new ones are tutor, checkout, plus and config.
  • Claude Haiku 5.5 runs through the Anthropic Messages API with a JSON schema for every reply, low effort, and one tool, create_bill.
  • PayPal Orders v2 creates and captures shop bills. PayPal Subscriptions v1 handles Plus. The PayPal JS SDK draws both buttons, including the card button.
  • Upstash Redis stores quotas, subscriptions and bills.
  • A script copies the 14 scenes from the page into the server, so the AI can only role-play scenes that exist on the site.

Where the money decisions live:

  • The model never sets a price. create_bill sends item names and quantities. The server looks each name up in a fixed price list and builds the PayPal order itself.
  • Only the server's capture marks a bill as paid, not the model and not the page. If the learner types the marker the server uses for "payment confirmed", it is stripped.
  • The server checks a subscription with PayPal before the device counts as Plus. It checks again once the last check is more than 6 hours old. One subscription covers at most 3 devices.
  • Cost has hard limits: 3 conversations of 20 turns per free device per day, 200 turns per Plus device per day, and 4,000 turns per day for the whole site.

Challenges I ran into

  • PayPal blocks payments between two Taiwan-registered accounts. My first sandbox merchant and buyer were both Taiwanese, and the subscription was refused. The demo merchant is now a US sandbox account.
  • The model kept correcting the same old sentence. A prompt rule was not enough. The server now drops any correction whose quoted words are not in the learner's latest message.
  • The shopkeeper sometimes just said the total. When the customer said "Eso es todo, ¿cuánto es?", the model answered with a number instead of calling create_bill, so no ticket and no button appeared. The prompt now says that the ticket only exists through the tool.

Accomplishments that I'm proud of

  • Both PayPal flows work end to end in the sandbox with the US demo merchant (tested October 22, 2026). The AI shopkeeper opened a bill, it was paid with a sandbox test card and captured, and the ticket got its PAGADO stamp. A Plus subscription paid with a test card lifted the daily limit; after cancelling, PayPal shows the subscription as CANCELLED.
  • The per-turn cost is measured, not guessed: one turn uses 1,700–2,900 input tokens and 60–400 output tokens, about US$0.0003–0.0004. From that, a full 20-turn conversation comes to roughly half a cent to one cent (estimated; later turns carry more history). NT$60 is about US$2, so a month of Plus covers roughly 200–400 full conversations (estimated). A learner who used all 200 turns every day would cost about as much as they pay; the daily cap is there for that case.

What I learned

  • Sandbox accounts have a country, and the country changes what PayPal allows. Test with the buyer country your users will have.
  • Keep money out of the model. The model decides when to open a bill; the server decides what it costs and whether it was paid.

What's next

  • A native speaker review of the lesson Spanish. It has not had one yet.
  • Find out what learners will pay. NT$60 a month is my estimate, not data.

Built With

Share this project:

Updates

Submission history