A fuel station network checks every tank daily, and opens an investigation when one drifts out of tolerance

Losses hidden in the daily noise

Each station recorded tank dips, pump readings and deliveries every day, on paper or in a spreadsheet. Head office saw those records only at month end.

That delay is where losses hide. A slow leak, or a meter that drifts a little each day, looks like ordinary noise on any single day. By the time the month’s figures are added up, the loss is real, but nobody can say which tank, which day or why. Looking into it means going back through weeks of paper and spreadsheets.

There was history to keep, too. Past stock records sat in each station’s own workbooks, and any new system had to bring them along rather than start from zero. The network needed the daily records in one place, checked as they arrive, with a clear record whenever a tank drifts out of line.

How we worked with them

We work as forward-deployed engineers, inside the team’s workflow rather than from a specification handed over a wall. We mapped the daily stock form line by line, and brought the stations’ past workbooks in, so no history was lost.

The owner decided the tricky cases, such as how to treat a gain compared with a loss, and those rulings are built into the checks. We built in short loops, and every change was tested on a safe copy of the data, never on a customer’s live account.

What we built

• A daily stock form for each tank and pump, plus deliveries. Pump readings are suggested from the day’s sales.
• Automatic variance checks. When a tank’s running variance passes its set tolerance, an investigation record opens on its own, with the reason written in.
• A monthly stock report with submission rates, out-of-tolerance tanks, open investigations, colour-coded loss bands, an evaporation view, a watchlist and a trend by area.

AI coding agents (e.g. Claude, OpenAI) build each change, with tests and an independent AI review on every change.

An investigation record that already says why

The system is in production. Stations record their readings every day, and a tank that drifts out of tolerance opens its own investigation record.

Each flag comes with an investigation record that already says why it opened, so “which tank, which day?” has an answer from the start. The monthly report then shows the wider picture: which stations submitted, which tanks are out of tolerance, what is still under investigation, and where losses fall in the colour-coded bands.

Pump readings are suggested from the day’s sales, so there is less to type on the daily form, and each tank’s readings and deliveries for the day sit in one record. Tolerances are settings, so a threshold can be adjusted without hard-coding it again for each station. The stations’ past workbooks were brought into the same system, so the history did not have to start again.