Data Tech 2026

Large language models are increasingly being used to write SQL, summarize datasets, and automate parts of the data science workflow. In some cases, they perform remarkably well. In others, they fail in subtle but systematic ways, including misinterpreting metrics, hallucinating joins, or reasoning inconsistently over schemas. This talk examines what actually happens when LLMs are applied to real structured data, using concrete examples to separate surface fluency from reliable analytical behavior.

We will explore a practical approach to improving reliability by providing LLMs with interpretable, structured components derived directly from the data itself. I will demo how exposing explicit statistical structure changes model behavior and reduces hallucination while improving reasoning over real datasets. The goal is to clarify where LLMs add value today, where they remain brittle, and which system design choices materially improve performance in analytics workflows.