This thesis investigates the use of large language models (LLMs) for conversational analysis of industrial telemetry data. The aim is to make complex operational data easier to access through natural-language questions while improving the reliability of numerical and structured reasoning. The proposed approach combines staged reasoning with external tools, where the LLM is used to interpret the user’s question, create a structured representation of the intended analysis, and generate executable computations. The numerical processing is then performed using deterministic backends such as SQL, pandas, and Polars, and the generated outputs are compared through an answer arbitration step. The work also investigates different reasoning sequences, retrieval strategies, validation mechanisms, and fine-tuning of an intermediate confirmation stage. Overall, the thesis focuses on how tool-augmented and structured LLM pipelines can be used to support more reliable conversational interfaces for industrial telemetry analysis.