PocketPause: one local AI card, then outside

This submission, titled "Touch Grass," presents the software "PocketPause," a small local app designed to provide one outdoor observation at a time, then disappear into the background. Users select a time (5, 10, or 15 minutes) and a general location like a street, terrace, or campus. The app uses an open-weight model to generate a few observation cues, which it turns into a single card for the user to read, save, and discard.

The app does not require an account, GPS, photographs, or any other input beyond the user's chosen location and time.

The app's operation involves a React frontend handling user selections, loading, and error states, while a small Node server validates requests and calls the local Ollama model. Qwen3, a non-thinking mode model, selects the cues. The model generates one exact JSON field with distinct values from a six-cue enum. The cue report shows that all cards generated were valid, though variety is limited by the finite vocabulary.

The fixed non-AI baseline also produces 12 usable cards, indicating that the AI addition does not significantly enhance the outdoor experience based on the measured evidence.

The app emphasizes local open-weight AI, generating locally without sending prompts to a hosted inference API. This keeps the activity private and makes the generation engine replaceable, although only Qwen3 has been measured here. The model weights are kept private, and the trade-off includes about 1.36 GB of weight data and resource-intensive generation on the user's laptop.

The app is licensed under MIT, and the model weights and upstream cards are licensed under Apache-2.0. The final product is a desktop software demonstration, not an outdoor trial.