Amazon (Global Engineering, Maintenance & Sustainability)

Agentic AI for BMS Trend Fault Detection & Insight Summaries in Building Operations

Facility operators often have limited time and tooling to interpret large volumes of building management system time‑series data and translate it into clear, actionable insights. This project developed an agentic AI prototype that ingested synthetic or anonymized HVAC and energy signals (such as zone temperature, discharge air temperature, setpoints, and compressor behavior) to detect trends and anomalies and produce plain‑English explanations and visuals. The system coordinated a repeatable loop of data retrieval, time‑series analysis (including trend decomposition, forecasting, change‑point detection, and comparisons of key parameters versus outside air temperature with basic weather/occupancy normalization), and an LLM-based narrative layer that generated “explain‑the‑why” summaries with guardrails and redaction. The prototype generates a weekly trend report with dashboard tiles from the synthetic dataset, supported by documentation such as model cards, a data dictionary, tests, and an evaluation summary describing errors, limitations, and next steps, while avoiding any PII or customer/site identifiers.