Offline Emergency Alerts Using Compressed Broadcast Codes and On-Device AI

Emergency alert systems face a major challenge during disasters: the same network problems and infrastructure failures that make communication necessary also make it harder to deliver messages. Existing systems like Cell Broadcast Service (CBS) are limited to short, fixed-language text messages, which makes it difficult to support people speaking different languages or provide instructions based on their specific needs without exceeding bandwidth limits. This paper proposes an offline-first system that solves this problem by combining a small, compressed cell-broadcast message with an on-device Small Language Model (SLM). The system can create localized and personalized emergency messages without needing internet access when the alert is received. The government sends a compressed message containing the disaster type, user role, evacuation route, and error-checking information within only 19–32 bytes, regardless of the message content. On the user's device, an asymmetric capability model adjusts the processing based on hardware. Tier-1 devices (below 4GB RAM) display a reliable predefined emergency message, while Tier-2 devices (4GB+ RAM with basic NPU/CPU support) use a local SLM to generate more detailed and personalized alerts. To make AI-generated messages safe, the system does not allow the SLM to create safety-critical information by itself. The AI only helps format and translate verified data into natural language. A rule-based validation system checks the output before showing it to the user. If the generated message does not match the verified information, it is rejected and replaced with a trusted predefined alert. This paper presents the first version of the system architecture and design. A working prototype is currently being developed, and testing is ongoing. Performance results and implementation details will be presented in future work.