DEV Community

Ayush Shekhar
Ayush Shekhar

Posted on

plot twist: i actually won a gemma 4 hackathon this time

so a couple months ago i wrote this post — "I didn't win the Gemma 4 challenge but here's what I actually learned."

well turns out the sequel is more fun. ScreenMind just took first prize (Open Track) at the Build with Gemma 4 hackathon on Kaggle.

i submitted it honestly not expecting much. same project. same idea. different hackathon. but this time it clicked.

what even changed?

honestly? not that much. the core features — MCP server, agent builder, meeting transcription, obsidian/notion sync — were all already there when i wrote the "didn't win" post. the project that "lost" and the project that "won" are basically the same project.

what i did ship between the two:

  • published to PyPI (pip install screenmind) — made it actually installable in one command
  • desktop launcher with system startup registration and graceful shutdown
  • wayland support for linux (grim + XDG Portal)
  • variant-aware Model Hub so you can swap gemma variants based on your GPU
  • pipeline parallelization and structured logging across the entire codebase
  • active monitor capture (beta)

basically i went from "cool demo" to "thing you can actually install and rely on." no new flashy features. just making the existing ones not break.

why i think it won

i think the judges liked that i wasn't stacking 5 different models together. most projects in this space do the obvious thing — whisper for audio, clip for vision, gpt for reasoning. three models, three inference pipelines, a bunch of glue code.

screenmind runs vision, audio transcription, and text reasoning on a single gemma 4 model. on 4GB VRAM. that wasn't some genius architectural decision — i literally just didn't have a good GPU and wanted everything to work on my laptop.

turns out "i was too broke for multiple models" is a valid engineering constraint that leads to interesting solutions lol.

the timeline that doesn't make sense to me

  • may: pushed first commit
  • june: didn't win the dev.to gemma challenge
  • june: hit hacker news front page anyway
  • june: people started covering it in chinese, japanese, korean
  • july: 190+ stars, 1800+ pypi downloads
  • august: won kaggle hackathon
  • october: presenting on github's open source friday livestream

i don't really have advice here. i just kept building the thing and stuff kept happening. if you shipped something and it "failed" — maybe give it another month before you move on.


repo: github.com/ayushh0110/ScreenMind
install: pip install screenmind

Top comments (0)