How we work: the systems we run every day

Problem

Prices were updated for several markets every week in a spreadsheet: turnover rate, discount level and currency conversion by hand, for each country separately. A mistake in one cell showed up in the store only when a customer noticed it.

Solution

A rules engine reads stock and sales history, calculates a turnover rate for each product, sets the discount level by the rules, and produces both a human-readable recommendation report and an import-ready file for the store platform, for every market in the same run.

Outcome

A price change across markets is one run, not a week of work. A person reads the report and approves it before the file goes to the store.

Problem

In a multilingual store a translation goes missing, languages drift apart or a product is left untranslated. Nobody notices, because the error is in one language among a large number of products.

Solution

A set of checks compares the live product feed and the site against the generation rules: translation gaps and consistency of coded values across languages. A second check makes sure the feed's structure matches what is actually exported from the store. The check returns a deviation report and says whether the feed passed.

Outcome

Silent errors are found before they appear in ads or in the store.

Problem

Building a search ad campaign by hand: keywords, ad copy, character limits, and making sure the ad doesn't promise anything the page doesn't say.

Solution

The generator reads only verified source texts, facts taken from the company's own pages, and the pricing rules, and produces an import-ready campaign. It refuses to write anything if a character limit is exceeded or a claim has no source.

Outcome

Campaigns are built faster, and not a single unfounded or over-length ad gets through.

Problem

Stock running out, products selling below cost and loss-making order lines had to be hunted down by hand in several reports, every day.

Solution

One view, prioritised by money at stake, combines stock, pricing and marketing signals into today's task list. A separate view flags products priced below cost.

Outcome

One list instead of searching for the same thing in five reports.

Problem

The same question got different answers from different reports. Some views had been written straight into a console with no version history. Management metrics required someone to run a query.

Solution

One set of versioned views, a table catalogue that documents itself, a guard that finds hand-edited views, and a scheduled run that writes the dashboard's data to a file with no server of its own. Every query has a cost cap.

Outcome

One truth for reporting logic, and a dashboard that is up to date without anyone starting anything.

Problem

When analysis is done with AI one session at a time, the correct result stays in the chat and the next session uses the old value. It happened to us: one session calculated the right cost, and every later one used the wrong one.

Solution

Every business figure is a row in a register: unit, validity period, source, status. Rules refer to the figure rather than containing it. Before anything is derived, it is looked up in the register; if the register says "disputed", work stops. The end-of-session routine refuses if a new figure appears without an entry.

Outcome

Figures stay current across several parallel sessions, and the same mistake doesn't repeat from one session to the next. This is the same discipline we train and deploy with clients.