This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
TrailTune is an offline bird call identifier for hikers and casual walkers. You open it at the trailhead, hit record, and it listens to birdsong around you and tells you which species it is hearing, with a confidence score and a one-line fact about each bird. It also keeps a small "life list" on your phone, so every walk becomes a mini scavenger hunt: "Can I hear 5 new birds today?"
The screen is meant to be the shortest part of the experience. You tap record, put the phone back in your pocket, and keep walking. The app buzzes gently when it detects a new species, so you look up at the trees instead of down at a display. At the end of the walk, you get a simple summary of the birds you heard and where.
It is for beginner birders, hikers, school nature clubs, and anyone who has stood under a tree thinking "what is that bird?" but had no mobile signal to look it up.
Demo
Video walkthrough: add link here (a 60-second clip recorded on a real trail with the phone in airplane mode).
Code
GitHub: add repo link here
How I Built It
- Model: An open-weight bird sound classification model (BirdNET-style, trained on open bird audio datasets), converted to a lightweight format that runs on a phone or laptop CPU. No GPU and no internet needed.
- Audio pipeline: The app records short 3-second audio windows, converts them to spectrograms, and feeds them to the model. Predictions are averaged across windows to reduce false positives from wind and noise.
- Local species filter: A small bundled list of species by region and season narrows the results to birds that are plausible for the area and month, which noticeably improves accuracy.
- Optional small open LLM: A tiny open-weight language model (running locally via llama.cpp) generates the short, friendly "bird fact" for each detection and the end-of-walk summary, so even the text part works with no connection.
- App layer: A simple progressive web app for the UI, with all data (life list, recordings, summaries) stored in the browser's local storage on the device.
- Workflow: Walk, record, get a haptic buzz on a new species, review the list at the end.
Why Does Open Innovation Matter?
- Works with no signal: Trails rarely have coverage. Because the model weights are open, the whole thing runs on-device. A closed cloud API would simply fail exactly where this app is most useful.
- Privacy: Audio recorded in a park or near homes never leaves the phone. No account, no upload, no server holding someone's location history.
- Free to run: There are no per-request fees, so a school club or community group can use it with 30 kids without worrying about a bill.
- Hackable: Open weights mean the model can be fine-tuned on local birds (for example, species common in a specific region of India) and swapped for a better one later. A closed API gives you no way to do that.
- Community-driven: Open datasets from bird recordists made the model possible, and anyone can contribute new regional recordings back.
My Agent Session
Add DevRelay session link here.
Prize Categories
- Best use of open-weight models
- Best local / offline inference project
- Best "Touch Grass" outdoors experience
Top comments (0)