School Bridge

Explains school messages in a parent's own language and drafts the reply in English.

Comment

Inspiration

Many immigrant parents receive school messages they cannot fully read. A missed deadline or a misunderstood permission slip affects their child. I wanted an assistant that explains the message in the parent's own language, lists what has to be done, and helps write a reply, while the parent stays in control.

What it does

School Bridge is an MCP server with a simulated voice-assistant demo. A parent speaks or types in Russian or Kazakh. The assistant:

  • explains a forwarded school message in that language
  • extracts tasks with deadlines and lists open tasks
  • marks tasks done
  • drafts an English reply with a back-translation, so the parent can check what it says

Nothing is ever sent to the school. State is saved in SQLite between sessions.

What is real and what is simulated: the Alexa+ experience is simulated by a local web page (browser speech recognition and synthesis). The MCP server (six tools, Streamable HTTP) is real, but the Alexa+ MCP Toolkit is partner-only, so I could not register it as an add-on or test it with a real Alexa+ client.

How I built it

Python, MCP Python SDK v1, SQLite and the Gemini API (gemini-flash-lite-latest).

  • An LLM router picks the tool for each request, but the code validates the choice against an allowlist. Unknown tools become "none".
  • All model output is validated in code. Untrusted text is length-limited and delimiter-stripped before it reaches the model.
  • The demo page listens on 127.0.0.1 only, accepts JSON only, checks the Host header and renders text with textContent.
  • 11 unit tests. Open source under the MIT license.

Challenges

  • Alexa+ MCP Toolkit access is partner-only, so I simulated the Alexa+ experience.
  • I checked the server's initialize handshake with protocol versions 2025-11-25 and 2025-03-26, but not against a real Alexa+ client.
  • Gemini's free tier sometimes timed out, so I added one retry and a longer timeout.
  • Kazakh output needed checking for mixed alphabets (a Latin letter inside a Cyrillic name).

Accomplishments

A working end-to-end flow in Russian, Kazakh and English with human-in-the-loop safety: the model never sends anything and cannot call tools outside the allowlist.

What I learned

That I can use agents to read and write drafts on my behalf and highlight tasks especially if there is several kids in the household. And it is easier to connect with tool with Alexa+ and receive messages and coordinate all activities easily.

What's next

Real Alexa+ testing once add-on access is available, per-customer state with account linking, a native review of the Kazakh phrases, and more languages.

Built With

Share this project:

Updates

Submission history