An agent surveys, builds and deploys the production management system at Pilo

Pilo produces and sells farmhouse cheeses. Its entire operation lived in a spreadsheet that one person filled in by hand.

The problem with a spreadsheet is not that it is useless: it is that it cannot say no. It accepts an impossible number as readily as a correct one, and the error only surfaces once a decision has been made on top of it.

And there were three concrete gaps the spreadsheet could not cover.

• Yield was fixed: a hand written constant, the same for every batch, that could not be checked against what each run actually produced.
• Sales were loose amounts: with no line detail, so there was no way to get from the product to the result.
• The margin column was empty: the figure that decides the business was the one nobody calculated.

And none of that is fixed with more careful data entry. It is fixed with a system that rejects what cannot be.

We built a web production management application, and what stands out is not only what it does: it is how it was made. The survey, the design, the code, the tests and the deployment were all carried out by an AI agent, with explicit human approval at every decision.

How it was built

• The survey started from the client real spreadsheet and was structured into 54 questions about how the business works.
• The agent designed the data model, wrote the code, built the tests and ran the deployment.
• No deployment went out without explicit human approval, and always with a dry run first.
• Scope and every business decision were approved one at a time, not in bulk.

Production traceability

• Milk intake with its price and the freight of each trip.
• Per batch production: from liters to units and kilos, with double weighing, at make and at the end of ripening.
• Real yield checked against theoretical yield, and shrinkage calculated rather than assumed.
• Per batch stock, so traceability runs from the batch to the sale.

Commercial and costs

• A catalog of cheeses and formats with retail and wholesale prices, for first and second quality, with history.
• Sales with line detail, accounts receivable and payments in pesos and dollars.
• Per run costs and monthly costs prorated by liters, which is what makes the real cost of each batch reachable.
• A dashboard by run and by period, and CSV export.

The validations the spreadsheet did not have

• It will not let you sell more stock than is available.
• It will not let you use more milk than came in.
• Without an exchange rate it saves nothing involving money.
• A sale that already has payments cannot be deleted.
• These are hard rules that reject the entry, not warnings that can be ignored.

Architecture and technology

• A web application in PHP 8 with a MySQL database, in production on managed hosting.
• 12 screens, 17 tables and 3,175 lines of code across PHP, SQL, JavaScript and CSS.
• An end to end test of 50 checks against a real server and a real database.

Learn more about artificial intelligence.

The business moved from a spreadsheet to a system in production, surveyed, built, tested and deployed within a single day.

Starting point

• The whole operation in a spreadsheet filled in by hand by one person.
• Yield fixed at 11.3% for every batch.
• Sales recorded as loose amounts, with no detail.
• An empty margin column.

What was delivered

• A 54 question survey over the client real spreadsheet.
• 3,175 lines of code, 12 screens, 17 tables and 2 migrations.
• An end to end test of 50 checks, with 0 failures, against a real server and database.
• Full traceability from batch to sale, with double weighing and real yield.
• Accounts receivable and payments in two currencies, and per run costs with monthly proration.
• A system in production, built and deployed within the day.

Why it matters

• Yield stopped being a hand written constant and became something measured batch by batch.
• And margin, which was the empty column, became calculable because sales have detail and costs have proration.
• The AI is not inside the product: it is in the fact that a complete management system was surveyed, built, tested and put into production in one day, with a person approving every decision.