AI data-quality agent

The problem

Master data in large organisations accumulates inconsistencies over years: duplicates, conflicting attributes, missing fields, and values drifting from an agreed standard. Generic null checks catch almost none of the issues that matter to the business.

This agent is one stage in a multi-agent pipeline for enterprise master data. Its job is to generate data-quality rules that reflect how the business actually uses the data.

Approach

The agent generates rules across six data-quality dimensions using business-grounded logic rather than structural checks alone. It runs on a FastAPI backend with a React front end, and the LLM layer is model-agnostic so the underlying model can be swapped without touching the pipeline.

The part worth writing about is the schema architecture. The first version had a hardcoded list of tables. That became an uploadable schema file that hot-reloads on the server, so the agent adapts to a new client instance without a code change or a restart.

What I took from it

Generating rules was not the interesting problem. Making the system adapt to a new client with no engineering work was. The most valuable time went into that boundary between configuration and code.