TouchGrass FieldAgent: An Open-Source Field Companion Designed for Zero Screen Time

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

The Problem

Modern outdoor and field apps suffer from a fundamental paradox: to navigate or record observations, they keep users glued to a glowing screen. Traditional LLM interfaces exacerbate this by returning lengthy conversational essays when all a user needs on the trail is an immediate, deterministic safety check and action plan.

Why Open Innovation Matters

Building this system around open-weight models (like Google Gemma) and deterministic agent control planes changes the operational paradigm:

1. Privacy & Security: Sensitive field notes and real-time location checkpoints remain off commercial ad-tracking servers.
2. Zero-Latency Output: Eliminates token bloat and conversational filler, ensuring quick execution.
3. Model Interchangeability: Standardizing prompt contracts allows effortless swapping between local Gemma runtimes and low-code edge control planes.

System Architecture & Workflow

The system utilizes Relevance AI as an explicit control plane orchestrating structured data transformation:

1. Input Parser: Ingests raw text or voice-to-text field notes containing location constraints and weather warnings.
2. Deterministic Guardrails: Strips conversational noise and enforces schema validation.
3. Inference & Execution (Gemma Engine): Processes the prompt under explicit JSON Mode constraints to guarantee structured key-value output (actionable_plan, safety_checklist, offline_summary).
4. Offline Dispatch: Emits a high-contrast summary card so the user can review the plan in 3 seconds and put their device away.

Live Output Schema Test

json
{
  "actionable_plan": [
    "Start now; use upper trail only and avoid lower ledge after 3:15 PM high tide.",
    "Limit route to a 20-minute out-and-back trail check on dry, stable sections.",
    "Turn around immediately at slick rock steps or wave splash zones."
  ],
  "safety_checklist": [
    "Watch footing on sea-spray-slick descent steps.",
    "Stay clear of edge exposure and lower ledge during high tide.",
    "Keep phone away; use audio/vibration only if needed."
  ],
  "offline_summary": "20-minute upper-trail check only at Sunset Cliffs; slick steps and 3:15 PM high tide make the lower ledge unsafe."
}

## Code & Repository

  

    

      
      
        rcortez056-spec
       / 
        touchgrass-field-agent
      
    
    

      
    
  
  

    

TouchGrass FieldAgent


An edge-optimized AI field agent designed to process raw trail observations, audio transcripts, and environmental conditions into structured, deterministic action plans.


Purpose


Built for the DEV Hacktoberfest 2026 "TouchGrass" challenge. The goal is to minimize screen time by delivering immediate safety checks and concise plans for outdoor activity.


System Architecture




Control Plane: Relevance AI Workflow Orchestration

Inference Layer: Google Gemma Open-Weight Model Engine

Data Format: Enforced Strict JSON Schema



Output Contract




{
  "actionable_plan": [
    "Start now; use upper trail only and avoid lower ledge after 3:15 PM high tide."
    "Limit route to a 20-minute out-and-back trail check on dry, stable sections.",
    "Turn around immediately at slick rock steps or wave splash zones."
  ],
  "safety_checklist": [
    "Watch footing on sea-spray-slick descent steps.",
    "Stay clear of edge exposure and lower ledge during high tide.",
    
…


    

    




  
  
View on GitHub