She asked for scripts.She needed a system.

How it worked

The pipeline is the boring part — that’s the point. Here’s how the recordings became a permanent site.

1. 1

   We listened to the problem under the request

   She asked for Python scripts. We asked: what would the scripts answer? Twelve questions, every day, mostly the same twelve. Scripts answer one question at a time. A system answers all of them, faster than people can ask. We scoped to the system.

2. 2

   Workbook walkthrough

   Thirty minutes of screen-share over her real workbooks. We mapped the schema — sheet purposes, cross-references, named ranges, the cost code structure, the meaning of every color and font convention. The walkthrough produced a schema doc that became the data contract.

3. 3

   ETL pipeline built

   A small Python service watches the workbook directory; on every save, it extracts the changed sheets into a clean analytical database. Every row preserves its source coordinates — file, sheet, cell range — so the audit trail follows the data into the cockpit.

4. 4

   KPI library defined

   The eight to twelve questions that drive 80% of her daily inquiries became named queries — codified once, available everywhere. Every KPI has a written definition, a calculation, and a list of source rows it depends on.

5. 5

   Cockpit and query interface built

   Static Next.js frontend, fast and offline-capable. Above the fold: the questions she gets asked. Below: the questions she asks herself, as a saved-query library. The natural-language input converts plain English to a structured query against the known schema — restricted, never hallucinating, always citing sources.

6. 6

   Refresh on save, no data team

   No nightly batch. No pipeline ops role. The workbooks remain her source of truth; the cockpit updates the moment she hits Cmd-S. We handed her the source code, the schema doc, and a one-page operating manual.