The Gemma 4 Good Hackathon

The Gemma 4 Good Hackathon · Main Track · Second prize

Trido

Built by Ardelyo and Azhar Ayatullah

A classroom whiteboard with Gemma 4 tools for mind maps, quizzes, and timers. The code supports online voice input and an Ollama text path; offline speech transcription is missing.

I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.

The idea worth stealing

Voice is the interface, not the accessibility mode.

The builders describe designing Trido with Pak Damar, a teacher in Bandung with disabilities affecting his hands and legs, and it never treats speech as a fallback. Every request becomes a function call that edits the canvas, the system prompt stays in English while the teacher speaks Indonesian, and a 2B model on Ollama supports text requests when the school’s internet drops; speech transcription does not work on that path. The builders report that animating a cursor to each AI action helped students follow the board.

What we checked

Date checked: .

Public repository at the submission revision, Kaggle writeup, and linked artifacts; code was read, not executed. Deployment checks and limits are described below.

Labels apply to each finding and the evidence described, not to the project as a whole.

Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.

  1. Eight tools, and the canvas is the API

    Verified

    The server defines eight functions, from add_mindmap_node to add_interactive_app, rejects calls to unknown tools or object IDs, and on the Ollama path parses fenced JSON when the 2B model skips native tool calls. Source

  2. Offline mode has no voice

    Inconsistent

    With Ollama, the transcribe endpoint returns an empty string, and the browser path is Chrome’s Web Speech API, which needs Google’s servers. When the internet drops, a teacher who cannot type is back to typing. Source

    What we need: A submitted build and network-disabled test showing local voice transcription, or a corrected offline-mode description.

  3. Gemma 4 on both paths

    Verified

    Online requests go to gemma-4-31b-it through the Gemini API, offline ones to gemma4:e2b on Ollama, chosen by probing both. The English system prompt tells the model to answer, and label nodes, in the teacher’s language. Source

Under the hood

A cursor with a speed limit
Each action first animates the cursor to its target with an eased move clamped between 350 and 1,500 milliseconds, then executes; actions drain from a queue one at a time, which the builders say helped students follow along.
The release is a start script
v1.0.0 ships no binaries. It is the source tree plus start.bat and start.sh, which install Node dependencies and open Chrome or Edge in --app mode, chosen because that mode exposes the Web Speech API.

Nuance

We read commit 564dc3b from May 17; the repository has 36 later commits adding mind-map, export, and attendance features. There are no tests, and we did not run the app. The classroom figures, six weeks and a 40 percent offline share, are the builders’. The demo video was recorded on Google AI Studio’s App Builder, which the writeup discloses.

Sources

Last updated: . Editorial updates do not imply a new technical check.

Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.

Builder? Add context, request a correction, or ask for a re-check →