Results came out today. ScreenMind didn't make it.

i'll be honest, i refreshed that page a few times hoping i misread it. i didn't. so okay, that's that.

then i did the thing you're supposed to do and actually read the winners instead of sulking. and i'm glad i did,because it answered a question i didn't know i had.

What won

go look at them, they're good:

• LIKAS, an offline disaster app for the philippines that keeps working when the cell towers go down
• AccessLens, basically eyes for blind and low-vision users, on-device
• a local postgres triage co-pilot someone built because HIPAA means they literally can't paste their db panic into chatgpt at 3am
• a phone agent that runs fully offline, peer to peer, so no prompt ever hits a server

every single one of these is a "here's a person who was stuck, here's how i unstuck them" story.

What i built

ScreenMind is a privacy-first take on microsoft recall. it watches your screen, understands what you're doing with gemma 4, and lets you search and chat with your own history later. "what did aachii say on discord" and it pulls up the actual message.
The engineering, i'm proud of it. it runs all three of gemma 4's modalities, vision audio and reasoning, on a 4gb gtx 1650. that constraint was the whole game. to make continuous analysis survivable on a card that small i had to build:

• a perceptual-hash cache so it doesn't re-analyze the same vs code window fifty times
• a chat that can cancel an in-flight analysis mid-inference and grab the GPU back in about a second, so you never wait a minute for a background job to finish
• meeting transcription straight off gemma's audio encoder, no whisper bolted on

honestly it's the most involved thing i've built.

The part that stung in a useful way

none of that is why the winners won.

i pitched ScreenMind as "recall but private." that's a feature comparison. it's me, an engineer, going "look how it works." the winners went "look who this helps." and that wins. every time, apparently.

i think i knew this in some abstract way but seeing it laid out next to my own submission made it concrete. the depth of the engineering didn't lose to better engineering. it lost to a better reason for the thing to exist.
that's a genuinely useful thing to get wrong once, early, on a challenge and not on something that mattered more.

So

i'm not shelving ScreenMind, it's open source and i'm still building on it: https://github.com/ayushh0110/ScreenMind

but the next thing i build, i'm starting from the person, not the spec. figure out who's actually stuck before i write a line of code.

congrats to everyone who won, and honestly to everyone who shipped something at all. that part's harder than it looks. On to the next one.