Utah Bird Project

behind the scenes

How it all works.

From birdsong to browser in under 30 seconds. Here's the full pipeline — every piece we built, wired, or configured.

the data pipeline

Bird sings

A bird calls or sings within range of the microphone.

BirdNET listens

The AI model analyzes 15-second audio windows and identifies the species with a confidence score.

Camera fires

Our Python script detects the new identification and triggers a 15-second video recording.

Data syncs

The sync service picks up the new detection and POSTs it to our FastAPI backend on Railway.

Backend processes

Deduplication groups detections into visits. New species trigger SMS alerts and eBird enrichment.

Dashboard updates

The website receives the new visit via Supabase real-time and it pops into the live feed.

the full stack

What we're using.

layer	technology	what it does
Edge compute	`Raspberry Pi 4`	Runs everything — BirdNET, camera, sync service. A $35 computer sitting in the backyard.
Audio AI	`BirdNET-Pi`	Cornell Lab's neural network identifies species from 15-second audio chunks. Runs locally on the Pi — no cloud needed.
Microphone	`Shure MV7`	USB cardioid mic. Picks up birds within ~30 feet. Originally bought for a podcast that never happened.
Camera	`Pi Camera Module 3`	Records a 15-second video clip every time BirdNET hears a bird. 1080p, no extra driver needed.
Sync service	`Python`	Watches BirdNET's local database, pushes new detections to the cloud every 10 seconds.
Database	`Supabase (Postgres)`	Stores every detection with timestamp, species, and confidence score. Real-time subscriptions for the live feed.
Backend API	`FastAPI`	Deduplication, enrichment, analytics. Turns raw detections into visits, species profiles, and leaderboards.
Frontend	`Next.js + React`	This website. Live feed, charts, species cards, build log. Deployed on Vercel.

The bird identification model (BirdNET) and species data (eBird) both come from Cornell Lab of Ornithology. Their open research makes projects like this possible. BirdNET can identify over 6,000 bird species worldwide and runs entirely on the Pi — no internet connection needed for identification.