Outside Cue - local AI that makes the screen the short part

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.

What I Built

An outside break doesn't need another feed. Outside Cue gives you three small
things to notice, then gets out of your way.

You describe your curiosity: the shapes of buildings, the sound around you,
or leaves and bark. A local sentence-embedding model matches those words to
an original catalog of twelve short observation cues. Pick an outdoor setting
and a short break duration. Review or print the cues, then start the timer.
You can finish early and optionally write down one detail you noticed.

The cues can be used from a comfortable seated outdoor spot. There is no map,
step target, leaderboard, notification stream, or need to photograph people.
The project does not determine whether a place or the weather is safe.

Demo

Watch or download the 20-second demo.

Captured-screen GIF tour.

The local 20-second MP4 demo uses actual browser captures of the form,
AI-selected cues, break timer, and reflection. Its banner explicitly labels
it as a capture tour, not real-time. It is not a field test or a recording of
someone completing an outdoor break. All 480 video frames decoded successfully.

Code

Outside Cue source and tests. Original application code and cue catalog are MIT licensed; model/runtime attribution is retained in the README.

How I Built It

The application is Node.js and plain browser JavaScript. Transformers.js 3.8.1
runs the Apache-2.0 MiniLM ONNX checkpoint locally on CPU. It creates normalized
sentence embeddings for the user's words and the cue catalog. A similarity
ranking selects three cues after applying the setting filter.

The model is essential to semantic matching, but it is not allowed to invent
instructions. All returned text comes from the reviewed catalog. Similarity
scores are not displayed as certainty about a user's needs.

The pinned model download is roughly 24 MB. A preparation command verifies
its metadata and ONNX hash. The running application disables remote model
loading. Missing files cause a clear startup failure rather than a hosted
fallback.

I used Codex for engineering and documentation assistance. Verification
includes 16 boundary tests and six real local model checks, all passing.
The six query checks are a small authored development set, not a claim of
general accuracy. I also tested the complete browser flow and mobile layouts
at 390px. I have not conducted a user study or outdoor field trial.

Why Does Open Innovation Matter?

A request as small as "help me notice architecture" should not require an
account or sending the thought to a remote model. After the one-time model
download, this project can compute the match locally without a hosted API key.

Open weights also make the boundary inspectable: the model selects from an
editable cue catalog. Changing the cues doesn't require a vendor service or
a prompt that might generate unsuitable activities. This is a narrow use of
AI, deliberately paired with ordinary deterministic filtering and a simple
screen that encourages leaving it behind.

Model attribution: Xenova/all-MiniLM-L6-v2,
an Apache-2.0 ONNX conversion of sentence-transformers/all-MiniLM-L6-v2.
Runtime: Transformers.js, Apache-2.0.

Only the overall category is intended. No partner technology prize is claimed.