Federated GraphQL integration of heterogeneous building sensor data for LLM-assisted utilization analysis in smart buildings

Abstract

Smart buildings generate extensive sensor data through wireless sensor networks (WSN), Internet of Things (IoT) devices, and machine-to-machine (M2M) communication, yet data fragmentation across heterogeneous systems prevents comprehensive utilization analysis. This work addresses the dual challenges of integrating disparate building data sources and making integrated data accessible to non-technical facility managers—a combination not previously evaluated in an operational real-world deployment. We developed a federated GraphQL architecture that unifies five heterogeneous systems at AAU Innovate, a multi-functional university innovation building: building management sensors (PIR, CO2, temperature), Microsoft 365 room calendars, entrance counters, coffee machine logs, and facility management records. The architecture maintains system independence while enabling cross-system queries through Apollo Federation [1], with each data source implemented as an independent subgraph. Five practical facility management questions were formulated with building operators, covering event attribution, resource allocation, external collaboration tracking, space utilization patterns, and policy intervention assessment. An expert user with full system knowledge successfully answered all five questions—establishing ground truth that cross-system data correlation enables insights impossible with isolated sensor networks—and providing a baseline against which LLM performance was evaluated. We then exposed the platform via Model Context Protocol and evaluated four open-source models (Qwen3, LLaMA 4, Mistral Large, Watt-tool) using a six-criterion scoring rubric across three to four independent attempts per question. Qwen3, the strongest model, achieved complete success on one of five questions and partial success on two, with failures attributable to context window overflow, state loss across multi-turn interactions, and multi-step reasoning breakdowns. These results suggest that LLM-assisted workflows under expert supervision, rather than fully autonomous AI agents, represent the viable near-term approach for building analytics applications.