I Built an AI That Identifies Birds—and Then Tells You to Put Your Phone Away 🌿
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.
What I Built
Imagine you're walking outside. You hear a bird singing somewhere nearby, but you don't know which bird it is.
You take out your phone, open an app, search for an answer, and start looking at your screen.
You identify the bird—but you might miss the experience that made you curious in the first place.
What if an AI bird identifier helped you discover the bird and then encouraged you to put your phone away?
That's the idea behind BirdWalk.
BirdWalk is an AI-powered bird-sound identification experience built around a simple philosophy:
Hear a bird. Discover something new. Then look up.
You can record audio, upload a supported audio clip, or select a bundled sample. BirdWalk sends the audio to a local FastAPI backend, where the real pretrained BirdNET model analyses it and returns its top predictions.
Instead of keeping users inside an endless identification interface, BirdWalk gives them a small nature-observation mission and a quiet phone-away timer.
The goal isn't to maximize screen time. It's to make the screen the shortest part of the experience.
What Makes BirdWalk Different
- Real AI inference: Uses the BirdNET-Analyzer and the BirdNET GLOBAL 6K V2.4 model
- Local processing: Audio analysis runs on the backend host rather than a third-party inference API
- Honest uncertainty: Model scores are presented as model scores, not as guaranteed accuracy
- Nature missions: Deterministic missions encourage users to observe their surroundings
- Phone-away experience: After identifying a bird, users can move away from the screen and spend time noticing nature
- Local journal: Completed experiences can be saved in the browser's local storage and deleted by the user
BirdWalk is designed for curious beginners, nature enthusiasts, students, and anyone who wants to learn more about birds without turning every outdoor moment into a screen-based activity.
Demo & Code
Try the product: https://birdwalk-three.vercel.app/
GitHub repository: https://github.com/sawanmani/BirdWalk
The repository includes the frontend, backend, BirdNET integration, sample metadata, model-output evidence, tests, model documentation, and third-party licence notices.
Built with:
- React 19 & TypeScript
- Vite & Tailwind CSS
- Python & FastAPI
- BirdNET-Analyzer
- TensorFlow-based model inference
The application is organized into a frontend, a local API backend, audio-processing services, bird-identification services, mission generation, and a local journal.
🎬 Video demonstration:
The video should show the real application: selecting or recording audio, running bird identification, viewing the result, completing a nature mission, and reaching the phone-away screen.
How I Built It
1. A Real Bird-Identification Model
BirdWalk uses the official BirdNET-Analyzer with the BirdNET GLOBAL 6K V2.4 model.
The backend validates and processes audio before passing it through the model's inference pipeline. It then aggregates the results and returns the leading species predictions. The shipped label file contains 6,522 labels.
Important distinction: A model score is not the same as real-world identification accuracy. The score describes the model's output, and it should not be interpreted as a guarantee that a bird has been correctly identified.
2. A Local Processing Pipeline
The frontend supports recording audio, uploading supported files, and selecting bundled samples.
The FastAPI backend validates and normalizes the audio before inference. Temporary audio files are cleaned up after processing.
The intended architecture keeps the inference path on the local backend host, avoiding a third-party cloud inference API for each identification request.
3. AI Should Know When Not to Be Confident
A bird identifier should not invent a convincing answer for every sound.
BirdWalk includes a no-reliable-detection path for cases where the output does not meet its detection criteria.
The repository's test report documents a synthetic birdless-noise test that returned no reliable detection. It also records a false positive from a sine-tone test.
That false positive matters. It's a reminder that AI outputs need to be communicated honestly, especially when people may mistake a numerical score for certainty.
4. Turning Identification Into an Outdoor Experience
After identification, BirdWalk generates a deterministic nature mission.
The current mission system uses predefined templates rather than a generative LLM. The same species maps consistently to a mission, making the behavior reproducible and testable.
The experience continues to a quiet phone-away timer and a local journal entry.
This is an intentional product decision: AI provides a clue, but the user should spend more time observing the world than interacting with the interface.
5. What the Tests Show
The committed pre-submission audit report documents:
- 39 passing backend tests, including three real-inference integration tests
- A successful frontend production build
- Five sample clips processed through the production pipeline
- Five top-prediction matches with the uploaders' claimed species labels
- A metadata-consistency test suite covering taxonomy, label counts, and evidence artifacts
These results are useful smoke-test evidence, not a scientific accuracy benchmark.
The five sample labels are source claims and have not been independently verified as ground truth. The observed agreement must not be presented as 100% model accuracy.
Why Does Open Innovation Matter?
Open innovation is central to BirdWalk because it lets the project use a real, inspectable bird-sound model rather than relying on a closed identification API.
Using open-source tooling and an available pretrained model made it possible to build an end-to-end experiment around audio processing, model inference, uncertainty handling, and an outdoor user experience.
It also made the implementation easier to inspect.
The repository records the model's observed outputs, documents its limitations, and separates the project's own code from third-party models and recordings.
That transparency matters because AI projects should be evaluated on what they actually do—not just on how impressive their demos look.
Local inference also gives developers more control over the processing pipeline and avoids sending each recording to an external inference API.
However, there are important qualifications:
- Local inference does not automatically mean that every part of an application works without internet access
- BirdWalk's full airplane-mode behavior remains unverified
- BirdWalk's original code uses the MIT licence, while the BirdNET model weights have separate CC BY-NC-SA terms
- The question of whether competition participation or prize money is compatible with the model's non-commercial terms remains unresolved
Open innovation makes experimentation possible, but responsible use still requires clear attribution, licence awareness, and honest reporting.
Prize Categories
BirdWalk uses BirdNET and open-source development tools, but the current implementation does not claim to use any of the challenge's listed partner technologies in a qualifying way.
The project can still be considered for the overall challenge award, subject to the official challenge rules and eligibility requirements.
What's Next?
BirdWalk has room to grow beyond identification. Possible next steps include:
- A personal collection of unique bird species
- A global leaderboard for bird discoveries
- Optional accounts and cloud synchronization
- Birdwatching challenges and achievement badges
- A privacy-conscious observation journal
- More extensive evaluation using independently verified recordings
- Physical-device microphone and offline testing
These are future possibilities, not features claimed as completed in this submission.
The core idea will remain the same: help people discover nature without encouraging them to spend more time staring at a screen.
BirdWalk — Hear one bird. Discover something new. Then look up. 🌿
Top comments (0)