This thesis investigates the design of a multimodal AI architecture that integrates sensor data streams, operational events, and technical documentation using embedding-based retrieval (vector databases) and an LLM-based agent. The agent follows a structured workflow that retrieves relevant evidence and executes a small set of constrained analysis tools (e.g., querying sensor/event data, summarizing episodes, and retrieving SOP excerpts) to generate traceable outputs. The goal is to evaluate an operational assistant that correlates sensor streams (e.g., MQTT/ROS2 data), process events, and domain knowledge (SOPs, maintenance logs) to support incident understanding and diagnosis.

The framework is realised as a working prototype and evaluated on public benchmarks and a simulated MQTT deployment. A pre-trained time-series foundation model (MOMENT-1-large) reaches a Recall@5 of 0.955 on the C-MAPSS retrieval benchmark (edging a hand-crafted statistical baseline and separating the operationally hardest degrading regime by 9.2 points), yet a per-channel Z-score remains the stronger detector of stealthy control-system attacks on the HAI dataset (AUROC 0.881 versus 0.600 for a foundation-model novelty score); the architecture accordingly uses the embedding for similarity-driven retrieval and rule-based statistics for detection. End-to-end, the constrained tool registry and citation policy gate produce grounded, citation-backed answers: on a systematic thirteen-query assessment spanning eight operator-query categories the agent reaches a correctness of 0.85 and a grounding rate of 0.89 with zero hallucinations (both hallucination-bait prompts are correctly declined against the telemetry), the residual failures concentrating on underspecified queries that draw a conservative refusal instead of a clarifying question.

The results contribute a reproducible blueprint for AI-assisted operational intelligence in industrial and robotic environments.