This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 1: Touch Grass.
What I Built
FieldQuest creates one small outdoor quest from two simple choices: how much time you have and what mood you are in. Read the prompt, pocket your phone, and go notice something.
It does not ask for your location. The quests are short, easy to adapt to a familiar public place, and designed to keep the screen out of the activity.
Demo
Code
How I Built It
FieldQuest uses Hugging Face Transformers.js to run the open-weight SmolLM2-360M-Instruct model locally in the browser. It uses a q4f16 ONNX export, with the model revision pinned in code. The quantized model weight file is about 272 MB; tokenizer and browser runtime assets add to the first download. After assets are cached, inference runs on-device and the app can work offline while that cache remains available.
The model suggests a short nature detail. FieldQuest puts it into fixed quest wording for the selected time and mood, then checks the phrase's format and a small blocked-term list. The list is only a limited guardrail. The app still asks people to choose a familiar public place, stay on paths, and use their own judgment about the conditions.
The app has no application server. Time and mood choices stay in page memory and are used only for local inference. No GPS, account, or analytics.
Why Does Open Innovation Matter?
The model weights and inference code can be inspected, changed, self-hosted, or swapped. Local inference means the two small preference choices do not need to go to a closed API, and generation does not need a connection while all model assets remain cached. Open weights make it possible to run this tiny tool without sending someone's mood or location to another service.
Prize Categories
Overall challenge entry. No partner category claimed.
Top comments (0)