UC Merced

Artificial intelligence has made remarkable progress in areas such as vision, language, and robotics, yet its deployment in real-world networked systems remains limited. These systems, ranging from wireless sensor networks to livestreaming platforms, present unique challenges such as unreliable data transmission, complex environmental dynamics, and strict system-wise constraints. This dissertation presents a suite of learning-enabled systems designed to tackle the unique challenges of real-world networked environments. It begins with a physically grounded propagation model for LoRa-based communication in orchards, enabling accurate signal estimation under dense canopy and terrain constraints. Building on this foundation, a model predictive control framework is developed for groundwater recharge optimization in agriculture, leveraging long-term forecasting and causal learning to ensure both water sustainability and crop safety. Inspired by the need for universal temporal reasoning, GrangerNet advances time series modeling by unifying Granger causality and delay differential equations to enhance interpretability and predictive accuracy. Extending this line of work into interactive networked applications, GenRTC addresses the real-time demands of generative AI by introducing a scalable system for low-latency visual captioning in live-streamed content. Together, these systems demonstrate how co-designing machine learning with domain and system knowledge enables robust, interpretable, and deployable intelligence in diverse networking scenarios.