This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
FloraTrail is a 100% offline, lightweight field assistant designed for hikers, gardeners, and outdoor enthusiasts exploring remote wilderness trails or tending off-grid gardens where cellular signal is completely non-existent.
When you're miles deep into a trail, cloud-based AI tools are useless. FloraTrail runs Google’s open-weight Gemma 2 model locally on your device via Ollama, connected to an intuitive Streamlit interface. It allows users to log field observations (leaf shapes, stem textures, environmental conditions), analyze plant safety and toxicity, and maintain a local log of trail notes without needing a single bar of cell reception or paying cloud API fees.
Demo
"Touch Grass" Test: Tested off-grid in an open outdoor area with Wi-Fi and Mobile Data completely turned off. FloraTrail processed local plant observations with sub-2-second response latency directly on a local laptop battery.
Code
📂 GitHub Repository:https://github.com/AbhasKorekar/floratrail-offline-ai..git
Clone and run locally (100% offline)
git clone https:https://github.com/AbhasKorekar/floratrail-offline-ai..git
cd floratrail-offline-ai
py -m streamlit run app.py
How I Built It
The app is architected with a strict Offline-First Stack:
AI Reasoning Engine: Google's Gemma 2 (gemma2:2b) running locally via Ollama. Gemma 2's high parameter efficiency allows it to deliver structured botanical analysis without requiring heavy GPU clusters.
Frontend Interface: Streamlit Python framework, leveraging st.session_state to maintain real-time field notes and environmental context filters (shade levels, soil moisture, proximity to water).
System Logic: Custom system prompt engineering enforcing Markdown outputs divided into three safety-focused field categories:
Identification & Context
Safety & Toxicity Check
Actionable Field Advice
Sample snippet from app.py executing local inference
response = ollama.chat(
model="gemma2:2b",
messages=[
{"role": "system", "content": "You are an expert outdoor botanist and wilderness guide..."},
{"role": "user", "content": f"Environment: {context_str}\nObservation: {user_observation}"}
]
)
Why Does Open Innovation Matter?
Open innovation isn't just an engineering preference—for wilderness and outdoor applications, it is an absolute necessity:
Off-Grid Reliability: Proprietary cloud models (like OpenAI or Anthropic) require active internet infrastructure. Open-weight models like Gemma 2 allow software to run in deep forests, mountain valleys, and rural farms where internet infrastructure doesn't exist.
Data & Location Privacy: Outdoor enthusiasts and foragers often keep secret trail coordinates or private garden locations. Keeping inference 100% local ensures zero personal or geographic data is harvested by cloud servers.
Zero Operational Cost: Outdoor utility tools should be free and accessible to everyone. Running open-weight models locally eliminates
subscription paywalls and per-token API metering.
My Agent Session
To ensure Gemma 2 delivered accurate, structured botanical advice without internet connectivity, I ran a multi-turn local prompt session via Ollama to calibrate responses.
Key Prompt Engineering Guidelines:
- System Persona: Enforced an outdoor botanist persona specialized in off-grid safety.
-
Structured Formatting: Configured strict Markdown outputs with 3 mandatory sections (
### 🌿 Identification & Context,### ⚠️ Safety & Toxicity Check,### 💡 Actionable Field Advice). - Safety Fallback: Instructed the model to explicitly flag hazardous plants (e.g., Poison Ivy, Giant Hogweed) whenever uncertainty exists in field descriptions.
Prize Categories
Prize Categories
Primary Category: Best Use of Gemma ($200)
Theme Alignment: Touch Grass (Hacktoberfest 2026 Week 1)
Overall Category: Hacktoberfest Open-Source AI Challenge Winner ($250)
Top comments (0)