Offline-First Multi-Agent Emergency Response with Privacy-Preserving Local LLM Coordination

The crisis

• Emergencies happen where connectivity fails — rural areas, disasters, infrastructure outages — exactly where cloud-dependent AI cannot reach
• The elderly, disabled, and those living alone are most at risk in emergencies and least likely to have reliable internet access
• 911 operators handle communication under extreme stress; AI-assisted coordination could reduce the burden on both sides of the call
• Medical records are sensitive — emergency AI that requires cloud upload creates a privacy barrier to adoption

About this research

Emergency response AI is almost entirely cloud-dependent, making it unavailable precisely when and where it is most needed. This work studies whether a coordinated multi-agent architecture — victim assistance agent and operator communication agent — running entirely on local small language models can deliver reliable, privacy-preserving emergency support. The nagents framework developed here enables structured tool calling and agent coordination on models like gemma3n that do not natively support it.

Key findings

• (In Progress)