AI/ML

AI-Powered Data Cleaning & Matching

Problem

The client's business operating hours existed as free text, entered by hand across thousands of records with no shared format or convention. Only about 15% of that data was usable as-is, which meant the operating-hours field was effectively unreliable for anything built on top of it. Cleaning thousands of inconsistent entries by hand wasn't a realistic option.

Approach

We built an AI normalization pipeline that read each free-text entry and converted it into a consistent, structured format, integrated with the client's proprietary systems through a custom MCP server. A second phase brought in external APIs to fill the gaps where the client's own systems had no data to draw on. Together, the two phases turned a scattered, inconsistent field into a dataset other systems could rely on.

"open Tuesdays after 1pm"   "MF 9-9"   ...thousands
             of inconsistent formats
                        |
                        v
              +------------------+
              | AI normalization |  <-- MCP --> client data
              | pipeline         |            systems
              +------------------+
                        |
                        v
            +-----------------------+
            | gap-fill via external |
            | APIs (phase 2)        |
            +-----------------------+
                        |
                        v
             uniform structured hours
               15% --> 96% accurate

fig. 1 — hours normalization pipeline

Outcome

The normalization pipeline transformed the reliability of that data.

Took business operating hours stored in wildly inconsistent formats — "open Tuesdays after 1pm", "MF 9-9", and thousands of other variations — and used AI to normalize them into a uniform structure. A second phase integrated external APIs to fill gaps where internal data was missing. Result: 96% accurate and consistent records, up from just 15%.

Have a similar problem?

Let's talk about what you're dealing with and how we'd approach it.

Start a conversation