Offline Emergency Alert Using Compressed Broadcast Codes & On-Device AI
An offline-first emergency alert system that combines a small, compressed cell-broadcast message with an on-device Small Language Model, so devices can generate localized and personalized alerts without needing internet access.
Existing systems like Cell Broadcast Service are limited to short, fixed-language messages, which makes it hard to support multiple languages or personalize instructions without exceeding bandwidth limits. This paper proposes a system where the government sends a compressed message (19–32 bytes) containing the disaster type, user role, evacuation route, and error-checking info. Based on the device’s hardware, it either shows a predefined alert (Tier-1) or uses an on-device SLM to generate a more detailed, personalized message (Tier-2). To keep this safe, the AI is not allowed to create safety-critical information on its own — it only formats and translates verified data, and a rule-based validation system rejects any output that doesn’t match.
