The Gemma 4 Good Hackathon
14 winning projects reviewed
An independent look at the fourteen projects that won Google DeepMind’s Gemma 4 hackathon, and the idea worth stealing from each.
I also participated in this hackathon with Memory Moment. This coverage was prepared independently after results were announced.
What we found
A few patterns from the winning projects we reviewed.
What AI did, and what stayed humanReading the submission revisions exposed useful boundaries: Project Rosie templates its synthesis specification, while Acuífero·Vigía computes alert levels without waiting for the model.
Public artifacts made parts of the work tangible. GEM-4’s checkpoints and Gilbeot’s adapter tooling were available, though simulator results and device performance were not independently reproduced.
Evaluation outputs mattered as much as model weights. PathOS’s saved notebook supported its 44% result; its 88% claim and TrueVoice’s headline results lacked saved outputs.
The shipped path sometimes narrowed the pitch: Trido’s offline mode lacked voice, KawanIsyarat loaded eight word-sign classes rather than sixteen, and MindBridge shipped the base model rather than its fine-tune.
Want this before winners are announced?
Finalist verification can run privately before awards are announced, giving organizers time to resolve open questions.
About the hackathon
A global online hackathon from Google DeepMind, run on Kaggle. Participants had seven weeks to build a solution to a real-world problem with the Gemma 4 open models, across five impact areas from health to safety and trust, and were encouraged to post-train the models, adapt them to a domain, and run them at the edge. A submission was a Kaggle writeup of up to 1,500 words with a public video under three minutes, a public code repository, and a live demo. Projects were judged primarily on the video, with the writeup and repository used to verify that the technology was real.
- Participants
- 2,234
- Submissions
- 1,608
- Tracks
- Main Track · Impact Track (Health & Sciences, Global Resilience, Future of Education, Digital Equity & Inclusivity, Safety & Trust) · Special Technology Track (Cactus, LiteRT, llama.cpp, Ollama, Unsloth)
- Prizes
- $200,000 in total: $50,000, $25,000, $15,000, and $10,000 for the four Main Track prizes, $10,000 for each of the five Impact Track areas, and $10,000 for each of the five Special Technology prizes. A project could win both a Main Track and a Special Technology prize.
- Requirements
- Kaggle writeup of up to 1,500 words; public YouTube video of three minutes or less; public code repository; live demo; a cover image in the media gallery.
- Judged on
- Impact and vision (40 points), Video pitch and storytelling (30 points), Technical depth and execution (30 points)
- Judges
- Glenn Cameron, Kristen Quan, Gusthema, Ian Ballantyne, Zhipeng Yan
From the official event page, checked .
The winners
-
GEM-4: Gemma Embodied 4 Physical Assistance
Takaki Maeda, Kazuma Hayakawa, Gayagayagaya4, Yuta_soyokaze, and Yushun
A wearable robot-arm research prototype using Gemma 4 E2B, trained on robot datasets and teleoperated episodes. Its first-person video pipeline estimates wrist motion; conversion into robot-training episodes remains incomplete.
Main Track
First prize
-
Trido
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.
Main Track
Second prize
-
PenguinAgent
And0Sha
A local video-analysis pipeline with SAM 3 penguin tracking, an 18-species classifier, and Gemma 4 report generation. We reviewed code and sample reports without running the tracker; literature grounding has gaps.
Main Track
Third prize
-
DEMENTOR
Inchara
An elder-care prototype with wearable sensor alerts and a structured Gemma triage response. The reviewed hub processes text and sensor data; the claimed speech transcription and vision paths are missing.
Main Track
Fourth prize
-
Project Rosie
Shashank Padala
A veterinary-oncology research prototype with code for ranking neoantigens and templating an mRNA synthesis specification. The public demo is synthetic; we did not run the tumor-to-construct pipeline or confirm its one-hour timing.
Impact Track
Health & Sciences prize
-
Acuífero·Vigía
Joaquín León Alderete, Lucas Ercolano, Joaquín Cadeiras, Fede Trejo, and Juan Andrés Quiroga Bonetto
A flood-warning prototype combining camera nodes, citizen reports, deterministic alert rules, and Gemma summaries. Code supports a templated SINAGIR export; performance figures are team-reported and no field pilot is documented.
Impact Track
Global Resilience prize
-
CodeBuddy
Adinda Panca Mochamad
A Python tutor prototype with photo-to-code, restricted execution, and prompts for five Indonesian languages. We reviewed the code without running it; spoken output uses a cloud service, and handwriting accuracy is unmeasured.
Impact Track
Future of Education prize
-
Gem-Care
Bilal Bounajma
A Gemma 4 speech-transcription fine-tune using structured clinical context. The repository records 19.6% word error rate under its evaluation method; the claimed baseline improvement and writeup’s 19.0% were not independently confirmed.
Impact Track
Digital Equity & Inclusivity prize
-
TrueVoice
Shiwon Oh and hyeminss11
A voice-deepfake research prototype with a frozen Gemma audio encoder and a 1.5 MB classification head. Its browser demo requires a notebook session; the claimed 5.2% equal error rate was not independently confirmed.
Impact Track
Safety & Trust prize
-
KawanIsyarat
Pradana Yahya
An Android communication prototype with on-device Gemma, a Whisper fallback, and BISINDO classifiers. The shipped word model has eight labels, despite a sixteen-class claim; we did not install the APK.
Special Technology Track
Cactus prize
-
Gilbeot
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.
Special Technology Track
LiteRT prize
-
PreVillage
Ashish Thapa
A navigator for Nepal’s government services with source retrieval in the public code. The claimed corpus size and private Gemma fine-tune could not be confirmed; voice transcription defaults to Gemini, and the live helpdesk timed out.
Special Technology Track
llama.cpp prize
-
PathOS
Dhairya Pandya, Shantanu Kharwar, and Rounak-agrwl
A histopathology research model with public Gemma weights and Ollama configuration. Training notebooks are available, but the claimed screening-to-report application pipeline and 88% accuracy were not independently confirmed.
Special Technology Track
Ollama prize
-
MindBridge
Huzayfah Patel
A Hindi mental-health screening prototype with a coded safety acknowledgment and on-device base Gemma after download. The fine-tune is separate from the app, and its claimed 25-point accuracy gain was not independently confirmed.
Special Technology Track
Unsloth prize