The Gemma 4 Good Hackathon

The Gemma 4 Good Hackathon · Special Technology Track · LiteRT prize

Gilbeot

Built by psymon

An on-device walking-assistant prototype with Gemma audio and vision code and a Korean-speech adapter. We reviewed the implementation without installing it; speech-accuracy and device-speed figures remain the builder’s measurements.

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

The idea worth stealing

The model localizes the arrow; code decides left or right.

Vision-language models are unreliable at left versus right, and a flipped word sends an elderly person the wrong way. Gilbeot asks Gemma 4 for the arrow’s tip and tail coordinates instead of a direction, then a comparison of two x-values settles it and overrides the sentence if the pixels disagree. The same refusal to trust a single path shows up in deployment: with no official route for an audio LoRA into a .litertlm bundle, the builder wrote a patcher that requantizes the merged weights on the original grid, and reports Korean character error rate falling from 13.1% to 5.0%; the evaluation set and script are unavailable.

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. Pixels override the word

    Verified

    If arrow_tip_x and arrow_tail_x differ by at least 0.05, the app derives left or right from them and rewrites the wrong word in Gemma’s sentence; if the sentence already agrees, nothing changes. Source

  2. The patcher and the sidecar are real

    Verified

    Three scripts requantize LoRA deltas on the original int4 grid; the audio adapter ships in the APK as an alpha-8 sidecar behind a C++ shim, and the grafted .litertlm is on Hugging Face. Source

  3. 5.00 percent CER

    Not independently verified

    The 134-utterance Korean set and the eval script the tools README names are not in the repository, so the drop from 13.14 percent is the builder’s measurement. Source

    What we need: The 134-utterance evaluation set, scoring script, and base-versus-adapter transcription outputs.

Under the hood

Forked flutter_gemma, four lines
Upstream never declares the MTP symbol the shipped library exports and rejects loraPath on .litertlm; the fork adds three MTP hooks and one LoRA path.
Judge build, real inference
The 200 MB demo APK cans the destination WAV, the T-Map route, and photos with GPS in their EXIF, but runs Gemma audio and vision on every one of them.

Nuance

We read the Dart, the patcher, and the reports; we installed nothing. Per photo the S23 needs about 12 seconds and the 2019 S10e about 50, which the builder reports rather than hides. The training set is Korea’s AI Hub elderly speech corpus, gated behind registration, and four unit tests cover the route service.

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.

The official submission and public profiles credit psymon. A full name was not confirmed.

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