Peter Yang demonstrates building a Japanese travel tutor using Gemini’s live-voice API, AI-assisted coding tools, and a spec with visual designs. This 18-minute tutorial is sponsored by Google.

The six-step build

  1. Explore the problem: ask AI to research before writing code and pose three clarifying questions. Yang chooses 100 travel phrases, organized into 10 lessons.
  2. Write a visual spec: pair requirements, APIs, and technical stack with phone and desktop screen mockups; edit the spec directly using a review tool.
  3. Build a first version: hand the spec to an agent; use Gemini Live for conversation and image generation for lesson illustrations.
  4. Test the actual voice loop: fix the wrong microphone selection, chat bubbles appearing all at once, and slow responses. Yang switches tools so an agent can use the browser to test the running app.
  5. Improve the lessons: teach phrases, then rehearse them in a simulated day in Tokyo. Iterate on the order, wording, and real usefulness of the exercises—not just the visuals.
  6. Launch: deploy to Vercel, resolve build issues, and gate the small personal beta with a passcode because voice API calls cost money.

What makes the demo useful

  • The coach first teaches each lesson’s 10 phrases, then role-plays situations where the learner says them aloud; the format can be adapted to other languages.
  • Yang’s spec contains a few representative screens, not merely a long requirements document; the pictures make design errors visible before implementation.
  • He keeps the Gemini API key out of the chat and places it in a local environment file. That still requires ordinary secret-handling and deployment discipline.
  • The image generator initially draws the coach as a rice ball; human review changes it to a character that fits the voice. The same scrutiny matters more for language pedagogy than for decoration.
  • Yang says the project took roughly three hours, about two spent iterating and testing. The first generated version is not the finished product.

“The point of this app is to actually teach you Japanese, not just to vibe code an app for fun.” — Peter Yang (14:46–14:49)