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)