Enabling Cloud-Level Accuracy in Edge AI through IoT Data Preprocessing

This paper presents a structured prompt construction framework that enhances the accuracy of local LLMs in interpreting IoT sensor data by preprocessing raw measurements into enriched textual representations. Evaluated on datasets from Raspberry Pi and various cities, results indicate that local model accuracy improved significantly, with indoor accuracy rising from 50.9% to 81.7% and outdoor from 63.7% to 89.3% when using enriched prompts. This approach addresses latency and performance issues in edge AI deployments, making it a valuable technique for practitioners seeking to optimize real-time analytics in smart environments.