This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Most modern AI applications are designed to maximize screen time, keeping your eyes glued to glass in endless conversation loops. TrailWhisper is designed to turn itself off.

TrailWhisper is a 100% offline edge nature companion and bioacoustic field guide powered by open-source AI. Built for hikers, trail runners, foragers, birders, and anyone experiencing screen fatigue, it puts nature first:

1. Ultra-Fast Capture (< 2 Seconds): Point your camera or snap an outdoor specimen (flora, fungi, aviary, pollinators).
2. On-Device Edge Vision: Open-weight Vision Transformers analyze the specimen locally on your device's GPU (WebGPU) in ~50ms without an internet connection.
3. Auditory Whisper (Eyes-Up Nature Experience): Instead of making you read paragraphs on a glowing screen, a natural voice speaks a concise naturalist insight directly into your earphones.
4. Tactile Sensory Challenges: Each sighting invites you to physically interact with the wild ("Touch the deeply furrowed white oak bark," "Smell the apricot aroma of chanterelle false gills").
5. Anti-Screen Auto-Sleep (Trail Mode): After speaking, a 6-second countdown bar immediately blanks the display into an ultra-dim, battery-saving OLED black mode with a breathing green leaf ring. Your eyes return to the canopy.
6. Smartwatch GPX Export: Your trail observations and GPS waypoints export directly into a .gpx file loadable into Garmin, Apple Watch, Strava, or AllTrails—allowing you to pack your phone away completely during hikes.

Demo

• GitHub Repository: https://github.com/Abhinav-Shrivas/TrailWhisper

How to Run Locally in 60 Seconds:

git clone https://github.com/Abhinav-Shrivas/TrailWhisper.git cd TrailWhisper npm install npm run dev 

Open http://localhost:5173 in any modern browser (Chrome, Edge, Brave).

Interactive Testing Highlights:

• Built-in Trail Presets: Test immediate edge AI classification without leaving your desk using curated specimens (Eastern White Oak, Fly Agaric Mushroom, Perched Songbird, Autumn Maple, Peat Moss).
• Live Camera Feed: Click Live Cam to aim your camera or webcam at houseplants or trees outside your window.
• Offline Trail Simulator: Toggle the TRAIL OFFLINE badge in the header to verify that the app runs with 0KB network traffic.
• AI Model Lab: Click the CPU chip icon in the header to run live hardware tensor benchmarks directly on your GPU.

Code

How I Built It

TrailWhisper is built from the ground up around open-source AI models and frameworks:

1. Google Gemma 2 (google/gemma-2-2b-it) — The Naturalist Reasoning Brain

We integrated Google DeepMind's open-weight Gemma 2 (google/gemma-2-2b-it) as our ecological reasoning and quest engine:

• Ecological Food Web Reasoning: When a hiker inspects a plant or mushroom, Gemma 2 synthesizes deep ecological connections (e.g. how white oak root exudates feed ectomycorrhizal networks, or how chickadee bioacoustic calls encode predator velocity).
• Dynamic Sensory Quest Generator: On the Sensory Quests tab, Gemma 2 generates contextual 3-to-5 minute tactile exploration challenges tailored to the immediate trail environment.
• Terminal Test Command: npm run test:gemma

2. Vision Transformer (ViT) — Hardware-Accelerated Edge Classifier

• Model: Xenova/vit-base-patch16-224 (open weights in ONNX format).
• Runtime: @huggingface/transformers (v3) running directly in the browser via WebGPU (Direct3D 12 / Vulkan) with WASM SIMD fallback.
• Zero Cloud Latency: Performs forward tensor passes in ~50ms without transmitting pixels across the internet.
• Terminal Test Command: npm run test:model

3. Offline Botanical & Wildlife Taxonomy Engine

• Pre-compiled taxonomic field knowledge for 50+ forest species pairing neural predictions with safety badges (Edible / Toxic / Protected), habitat notes, and tactile field cues.

4. Anti-Screen UI & Local Field Journal

• Built with Vite, React 19, and a custom Vanilla CSS design system styled with deep forest obsidian tones (#07100b) and bioluminescent accents (#4ade80).
• Observations, timestamps, and GPS coordinates are stored exclusively in local IndexedDB/localStorage with 0 external telemetry.

Why Does Open Innovation Matter?

1. Datacenters Can't Hike With You

Frontier models (GPT-4o, Gemini Cloud) reside in multi-billion dollar datacenters. When you hike into a valley, canyon, or national park, cellular reception vanishes—and cloud AI goes completely dark. If an assistant requires 5 bars of 5G to tell you whether a mushroom is toxic, it fails the outdoors. By running open weights locally on device via Transformers.js and WebGPU, TrailWhisper operates reliably in Airplane Mode with 0 bytes sent over the wire.

2. 100% Location & Foraging Privacy

Trail paths and foraging spots are personal. Commercial closed APIs log GPS waypoints and store photos on external servers. Open-source AI ensures that all coordinates, images, and notes remain strictly on the user's hardware.

3. $0.00 Operating Cost Forever

No subscription fees, no credit card paywalls, and no metered API billing.

4. Freedom to Build an "Anti-Screen" App

Commercial AI companies optimize their products for continuous user engagement and endless chat loops. Open-source AI gave us the autonomy to build an AI that does the opposite: answers in 2 seconds, speaks aloud, and intentionally turns off the screen so you can touch grass.

My Agent Session

This project was built and architected in pair-programming collaboration with Antigravity (Google DeepMind's agentic AI assistant). The agent assisted with:

• Designing the anti-screen WebGPU edge pipeline and auto-sleep countdown mechanics.
• Integrating Google's Gemma 2 open-weight reasoning model.
• Validating model inference with automated standalone Node scripts.
• Configuring GitHub Actions CI/CD to ensure continuous model testing on every commit.

Prize Categories

1. Best Use of Gemma ($200 USD)

• Model Used: Google Gemma 2 (google/gemma-2-2b-it) by Google DeepMind.
• Implementation: Gemma 2 powers the ecological reasoning engine and the real-time dynamic quest generator, explaining complex forest symbiotic relationships and creating sensory outdoor missions without cloud dependencies.
• Verification: Run npm run test:gemma to test the open-weight reasoning pipeline.

2. Best Use of GitHub ($100 USD)

• Continuous Integration: Implemented an automated GitHub Actions workflow (.github/workflows/ci.yml).
• Automated AI Testing: On every push and PR, GitHub Actions automatically spins up an environment, runs npm run test:gemma to verify the Gemma 2 reasoning engine, and compiles the production bundle with npm run build to prevent regressions for open-source contributors.