TrailSense - Google DeepMind x Cactus Compute Global Hackathon
AI Tinkerers - San Francisco
Hackathon Showcase Finalist

TrailSense

Team led by a Nutanix Senior PM and ex-Intel/AMD engineer with 20 years of experience, specializing in full-stack Next.js, Python, and agentic AI integrations.

1 member Watch Demo

TrailSense: Offline Voice-Guided Safety Agent for the Visually Impaired/Lost Hikers and Disaster Relief/Emergency Workers

TrailSense is an offline-first, voice-controlled navigation and safety assistant designed for visually impaired individuals, senior citizens, and hikers in low-connectivity environments.It also aids the emergency workers like wildfire relief/earthquake relief/EMT/paramedics.

It intelligently decides where computation should occur, locally on-device or in the cloud based on context, confidence, safety risk, latency, and privacy. User can also choose to execute on device using SLMs (Small Language Models) or in cloud with LLMs ( Large Language Models)

⚙️ Architecture

TrailSense combines:
• FunctionGemma for structured tool-calling and decision-making
• Cactus Compute as the local runtime for fast, private execution
• Gemini APIs as a cloud fallback for complex reasoning and contingency planning
• A lightweight offline trail graph for routing without internet

♿ Accessibility-First UX

Designed specifically for visually impaired users:
• Voice-first interaction
• Large accessible buttons
• High-contrast mode
• Spoken step-by-step directions
• Audio-optimized navigation language
• Offline operation

No map interface required.

🔄 Hybrid Inference Strategy

  1. Always attempt local inference first
  2. Use structured tool calls for route computation and hazard detection
  3. Evaluate confidence + safety criticality
  4. Escalate to Gemini only when:
•	Confidence is low  
•	Situation is safety-critical  
•	User explicitly requests enhanced reasoning  

This maximizes:
• Privacy
• Latency performance
• Edge/cloud ratio
• Cost efficiency

🗺 How Navigation Works (Without Google Maps Offline)

TrailSense does not rely on Google Maps offline.

Instead, it uses:
• Pre-downloaded offline trail graphs
• Local GPS positioning
• On-device shortest-path routing
• Structured safety tool calls
• Text-to-speech for spoken guidance

This ensures reliable navigation even with zero connectivity.

🔐 Privacy by Design
• Local-first inference
• Minimal data exfiltration
• Cloud usage only when needed
• User-controlled “Cloud Allowed” toggle

🏔 Real-World Impact

TrailSense enables:
• Blind hikers to navigate safely
• Seniors to walk independently
• Outdoor explorers to recover from disorientation
• Emergency guidance when network coverage fails

AI Tinkerers Cactus Compute Cursor Gemini API Google Google DeepMind

TrailSense: Offline Voice-Guided Safety Agent for the Visually Impaired/Lost Hikers and Disaster Relief/Emergency Workers

Summarizing URL...
Summarizing URL...