The project champions open innovation and offline AI as essential for usable, private, region-specific wildlife identification, enabling model inspection and reliability without cloud dependence.

Its offline approach emphasizes privacy, zero per-request costs, verifiability, and MIT-licensed analyzer and own-code components, with BirdNET weights marked as non-commercial.

Privacy and control are central: quest preferences are processed locally on the user’s computer, avoiding remote AI generation and limiting data exposure.

BirdNET analyzes about ten seconds of external audio to identify likely bird species, with results feeding Gemma to generate concise field-journal notes limited to the detected species.

The app also identifies plants, birds, and insects from photos, offers nearby observation cues, warns of local dangers, and creates short field missions to encourage exploration.

Presets target Alberta locations—Inglewood Bird Sanctuary, Fish Creek, Frank Lake, Kananaskis, Elk Island, Waterton—and the app supports phone voice memo, laptop view, or a Raspberry Pi for field use.

Core workflow runs through a browser-based microphone input, a Flask server, BirdNET 2.4 for ID, an optional local LLM via Ollama for notes, and a local SQLite log with de-duplication.

A demo is available via CI with a README and a quick BirdNET sample test; the author details converting BirdNET to ONNX and the challenges overcome.

Technical specs include 8 kHz recording, 8-bit μ-law, about five minutes of storage per note, and notes up to 30 seconds; weather prompts are guided by SerpAPI integration.

On-trail workflow features passive edge listening, sub-30ms on-device detection, spoken field notes, offline journaling, and an auto-pocket OLED sleep mode to conserve battery.

The field guide covers 6,629 offline species entries, each with a name, short summary, and a Gemma-generated fun fact grounded to Wikipedia content, checked for source reliability and kid-friendliness.

Three modular bricks—Web UI, Gemma LLM, and audio classification—run via Arduino App Lab, with Edge Impulse retrained birdsong models and local samples.