This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.
What I Built
I built Trail Buddy, a small web app for people who want to notice more on a walk. Describe something you saw or heard—a bird call, a changing leaf, a patch of moss—and it finds the closest match in a pocket-sized nature guide. It also offers simple prompts to encourage looking, listening, and exploring outdoors.
The app is designed to make the screen a quick stop: write a note, get a starting point, and look up at the world again. Matches are suggestions, not definitive species identifications.
Demo
Try the live app: Trail Buddy
Code
Explore the source: Trail Buddy on GitHub
How I Built It
Trail Buddy is built with HTML, CSS, and JavaScript. It uses Transformers.js to run the open-weight MiniLM sentence embedding model in the browser.
When someone submits an observation, the model turns it and each field-guide note into an embedding. The app compares them with cosine similarity and shows the closest match. The journal is stored in the browser’s local storage.
The first model load needs an internet connection to download the library and model. After that, inference happens in the browser; this version doesn’t guarantee fully offline use.
Why Does Open Innovation Matter?
Using an open-weight model lets Trail Buddy run inference in the visitor’s browser without sending their observation to a hosted AI API. That means no inference API key, no app server handling personal field notes, and no per-request model bill.
The approach is also easy to inspect and adapt. I can change the field-guide notes, try another embedding model, or adjust the matching logic without depending on a closed model provider. For this project, that flexibility and local processing fit the goal: a simple, private nudge to go explore.
Prize Categories
I’m not entering a partner category; Trail Buddy doesn’t currently use the listed partner technologies.
Top comments (0)