Leveraging AI-Powered Small Language Models for Real-Time Disaster Communication and Response Optimization

This paper investigates the use of SLMs in automated response planning and in real-time communication during disasters in a scenario where there is no extreme bandwidth and communication is scarce. The need to alert the population in the case of a disaster is pointed out. The article establishes the relevance of the topic: the growing frequency and scale of disasters render the speed and reliability of alerting systems critically important, whereas large cloud-hosted LLMs are impractical due to their substantial bandwidth and energy requirements. The objective of this study is to assess the feasibility and operational value of SLMs within post-disaster communication networks and to formulate governance-informed implementation practices for their deployment. The architectural and empirical work on the model and its prototype is the novel aspect of the research. The novelty of this work lies in a systematic comparison of architectures and prototype validation: a review of the literature together with experimental case studies demonstrates the feasibility of local SLM inference (Llama-3 8B, Qwen-2.5 7B) on single-board accelerators (Jetson Orin AGX) with INT4 quantization and parameter-efficient fine-tuning (LoRA/LoRI). The research spans fields such as power and usage latency, document semantic trust normalization, misinformation detection, hybrid BLE–LoRa networking, and Delay-Tolerant Store-and-Forward routing. The assessment indicates that for primary response purposes, SLMs can be used with the level of accuracy needed in the first hour of response at practically zero cost and therefore can be utilized in the first response hour. This logic will prove helpful to AI practitioners solving operational problems in assistance and rescue, architects of emergency communication systems, and disaster planners.