What I Built

Aurelia is a digital-detox game for Android that is boring on purpose.
The whole interface is black-and-white terminal text: ASCII progress
bars, a pixel butterfly, oversized level numbers. The color lives
outside, which is the point.

It gives you quests that can't be finished on the couch. Spot a bird.
Photograph a mushroom. Find the strangest texture within a hundred
steps. You take a photo as evidence, and a neural network running
entirely on the phone checks what is actually in the frame. Verified
finds pay XP; at 440 XP your companion reaches IMAGO, its final form.
Completed quests fill a photo journal, the only part of the app allowed
to show color.

It's for people whose screen time embarrasses them a little (mine does),
and for anyone tired of apps that claim AI verified something it
obviously couldn't. Aurelia would rather admit its limits than fake
certainty.

Code

mykeyy / aurelia

AURELIA

A boring app for a more interesting life.

Aurelia is an offline digital-detox game for Android. It hands you quests that can only be finished away from the screen: spot a bird, photograph a mushroom, leave the phone alone for ten minutes. You come back with a photo as evidence, an on-device neural network checks what is actually in the frame, and a pixel butterfly on your home screen grows from EGG to IMAGO as the XP adds up.

The interface stays black and white the whole time. The color is outside.

Built for the Hacktoberfest 2026 Open-Source AI Challenge, Week 1: Touch Grass. Source: https://github.com/mykeyy/aurelia

70-second demo video
Signed APK — v0.1.0 (arm64, 56MB; universal build included)

The 60-second tour

Twelve quests on the board. Ten are photo-verified discoveries (flower butterfly, bird, mushroom, dog, insect, squirrel, landmark, viewpoint street furniture). One is a timed outdoor break and one…

View on GitHub

How I Built It

I described the app I wanted and Devin, Cognition's software agent,
built it end to end while I steered. Two independent critic sessions
then reviewed the code cold and scored it 7.4 and 7.1 out of 10. The
defects they reported — a duplicate-XP race, evolution stages that were
mathematically unreachable, a break timer you could forge by changing
the clock, a self-cancelling inference effect — are fixed, and
QUALITY_REPORT.md records what was verified on a real emulator and what
wasn't.

The AI is real and small: MobileNetV2-12, a 13.9 MB Apache-2.0 model
from the ONNX model zoo, bundled inside the APK and executed by ONNX
Runtime Android. The app never requests the INTERNET permission, so
inference is offline by construction. Each photo quest owns a set of
ImageNet-1000 classes; a photo earns MATCH when the summed softmax
probability mass on that set reaches 0.25, UNCERTAIN between 0.08 and
0.25 (you see the raw top-5 predictions and decide for yourself), and
NO_MATCH below 0.08.

I chose summed probability mass over top-1 matching because top-1 lies
sometimes. The model once called a misty forest photo "jellyfish" at
100% confidence, but the mass on the bird classes was about 2×10⁻⁷, so
the quest still correctly returned NO_MATCH. The app also refuses to
fake certainty where the model has none: break quests log a timer and a
photo while admitting the photo can't prove you were outside, and
subjective quests go to the journal with no verdict at all.

The rest is ordinary Android done carefully: Kotlin, Jetpack Compose,
Room + KSP for persistence, CameraX for capture, the system photo
picker for imports. The design constraint is monochrome until the real
world provides the color.

Why Does Open Innovation Matter?

This app exists because the model's weights are on my disk. A photo
classifier that has to work in airplane mode cannot be a closed API
call, and the places Aurelia wants you to go — ridgelines, the far end
of the park — are exactly where there is nothing to call anyway. Open
weights are the only reason a no-network-permission app can carry a
neural network.

The license mattered in practice, too. Apache-2.0 lets me bundle the
13.9 MB file inside the APK and redistribute it, with no key server and
no usage quota for my users.

The biggest effect, though, was on the design. Because the model is
open and inspectable, I could run it on my own photos, watch it say
"jellyfish" about a forest, study the real probability distributions,
and design around its failure modes instead of writing marketing over
them. The UNCERTAIN verdict exists because the model was open enough to
doubt. So do the always-visible top-5 predictions and the quest copy
that admits a photo can't prove you went outside. A score endpoint
behind an API would have dared me to trust it, and the honest version
of this app would have been unbuildable.

DEV Community

Aurelia