This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
Modern hiking and outdoor apps suffer from a tragic paradox: they encourage you to stare down at a glowing screen while walking through ancient forests and alpine peaks. Turn-by-turn navigation chimes buzz on your wrist every 30 seconds. Hikers miss the rustle of leaves, the scent of damp hemlock needles, and the warning shift of mountain winds.
Even worse, the moment you hike a mile past the trailhead, cell towers vanish. The second you enter deep wilderness, every proprietary cloud AI assistant (ChatGPT, Claude, Gemini API) turns into a completely useless spinning wheel. If a hiker sprains an ankle, gets stung, or loses daylight, closed AI leaves them stranded.
I built TerraGemma: an offline-first backcountry expedition companion powered by Google's open-weight Gemma 3 running 100% locally on-device.
TerraGemma is engineered around a radical design principle: the screen must be the shortest part of your outdoor experience.
Instead of keeping you glued to a feed or turn-by-turn phone map:
Selecting a backcountry trail calculates astronomical solar timestamps (Sunrise, Sunset, Civil Dusk) using offline NOAA formulas and applies Naismith’s Rule with elevation gain penalties. A color-coded daylight status badge and a hard Turnaround Time ensure hikers never get caught on the mountain after dark, displayed alongside an interactive OpenStreetMap route.
Gemma 3 synthesizes memorable physical sensory anchors (the sound of feeder streams, shifting ridge winds, distinctive boulder cairns) so hikers navigate with their natural senses instead of checking turn-by-turn phone notifications. Clicking "Listen & Pocket Phone" activates the offline Web Speech API to read the briefing aloud, allowing the hiker to stow their device and hike hands-free.
In deep backcountry with zero cellular signal, emergency medical decisions cannot wait. Gemma 3 evaluates acute symptoms offline—providing immediate 3-step triage actions, critical contraindications (what NOT to do), evacuation urgency, and international 3-blast whistle rescue signaling protocols.
An offline botanical and wildlife field guide comparing dangerous lookalikes side-by-side (Poison Ivy vs. Virginia Creeper, Queen Anne’s Lace vs. Deadly Poison Hemlock, Black Bear vs. Grizzly) alongside mindful "Touch Grass" missions designed to foster genuine connection with the forest.
Here is how TerraGemma works in practice at the trailhead:
The complete source code is open-source under the MIT License on GitHub:
👉 https://github.com/shivamgravity/terragemma
TerraGemma is built in Python using Streamlit, engineered specifically for 100% offline, zero-cloud execution on consumer hardware:
gemma3:1b (Primary Edge Model):gemma3:4b (Analytical Model):keep_alive: 60m to ensure warm model weights stay in memory.window.speechSynthesis. Synthesizes natural speech locally in the user's browser without calling external cloud text-to-speech APIs.folium & streamlit-folium):
flowchart TD
A["Hiker at Backcountry Trailhead (Airplane Mode / Zero Signal)"] --> B["TerraGemma App (Streamlit Local Engine)"]
subgraph "Deterministic Outdoor Engines (Zero Dependencies)"
B --> C["NOAA Solar Engine (Sunset, Twilight, Golden Hour)"]
B --> D["Naismith's Rule Engine (Turnaround Time & Margin)"]
B --> E["OpenStreetMap Visualizer (Waypoints & Elevation)"]
end
subgraph "Open Edge AI Engine"
B --> F["Google Gemma 3 (1B / 4B via Local Ollama)"]
F --> G["Sensory Landmark Generator ('Eyes-Up Guide')"]
F --> H["Wilderness First Responder (WFR Triage Matrix)"]
F --> I["Botanical & Toxic Lookalike Field Assistant"]
end
G --> J["Offline Web Speech API Synthesizer"]
J --> K["🔊 Audio Briefing: 'Listen & Pocket Phone'"]
K --> L["Hiker Steps Onto Trail with Eyes Up & Mind Clear 🌲"]
In a coffee shop, relying on closed cloud AI APIs (like OpenAI or proprietary cloud endpoints) is convenient. In the wilderness, a closed API is 100% dead.
When you step past the trailhead into a national park, canyon, or alpine pass, cell signal drops to 0 bars. Closed APIs become infinite spinners. Open-weight models like Gemma 3 run natively on edge devices. You carry a world-class naturalist and wilderness first responder with you without a single cell tower in sight.
Commercial trail apps frequently package high-precision GPS tracks, departure times, and personal health metrics to sell to ad brokers and data brokers. TerraGemma runs 100% locally: your coordinates, trail notes, and emergency symptoms never leave your device.
Closed APIs charge meter fees per thousand tokens and enforce rate limits. On a multi-day backpacking trek, an open-weight model costs $0.00, will never hit an API rate limit, and operates completely indefinitely.
Google's Gemma 3 (1B) requires under 1 GB of memory. It runs on consumer laptops and edge hardware in hundreds of milliseconds without draining critical battery reserves needed for survival navigation.
I built and verified TerraGemma using DevRelay: