Abstract: Rural roads account for a disproportionate share of road fatalities, underscoring the urgent need for intelligent, scalable, and infrastructure-compatible safety solutions. This project advances rural road safety through a unified vision-language framework designed for comprehensive rural highway scene understanding. Leveraging video streams and multimodal data, the framework performs integrated weather analysis, pavement wetness assessment, traffic congestion detection, and special-object recognition (e.g., animal crossings), followed by automated, context-aware safety alert generation. To enhance interpretability and operational decision support, the system employs a multi-agent vision-language architecture in which specialized expert models generate structured scene representations and perform contextual risk reasoning. The distilled knowledge from these expert agents is transferred into a lightweight, edge-deployable model, enabling real-time inference under resource-constrained rural deployment settings. Experimental results demonstrate that multi-agent vision-language reasoning effectively supports weather, pavement, and congestion understanding while producing semantically faithful, risk-aware outputs. Overall, the proposed framework offers a scalable pathway toward next-generation rural transportation warning systems.