A workshop reference — and a working showpiece

01Data organisation and structure

Most businesses already own the data they need. It is just scattered across PDFs, catalogs, price lists and somebody's notebook. We turn it into one clean record per thing — a record a person can read and a program can calculate with.

One layout for everything. Here the raw material was manufacturer data sheets, a printed catalog, dealer tables and my own workshop notes. For each of the 545 products they became one record with the same layout: the name in Russian and English, what the product is, what it is mixed with and in what proportion, how it is applied, how long it dries, how it is packed.

Classification. Every product is sorted two ways: by chemistry (acrylic, polyurethane, polyester, water-based, oil and so on) and by the job it does (primer, lacquer, hardener, thinner). That sorting is what produces the 16 category pages and the filters. Nobody maintains those lists by hand.

Every value says where it came from. A number without a source is a rumour. Each value in a record carries a tag: the manufacturer's site, a data sheet, a dealer, the catalog, a finishing schedule, workshop notes — or “recommended”, which marks our own engineering estimate and carries the reasoning behind it. You can see these tags next to the numbers on any product page.

When sources disagree. They often do. We do not quietly pick a winner. The record keeps each statement with its source, and where two sources give different ratios for the same hardener or thinner, the calculator offers them as separate choices labelled by source, with the manufacturer's own figure first.

Nothing is overwritten. Every change to a product is saved as a new numbered version with the date, the author and a note. Earlier versions are never edited, so any past state can be opened and brought back — and bringing it back is itself recorded as a new version.

02AI integration

AI is good at reading and at explaining. It should not be trusted blindly. We use it for the first two and build the checks that the third requires.

Reading documents. A data sheet — a PDF or a photo of a printed page — is handed to an AI model with strict instructions: copy every number exactly as printed, list what the sheet does not say, never invent a figure or a product code. The retyping is done by the machine. What comes back is a draft, not a publication.

A person stays in charge. The draft is laid beside the current record, field by field, with the differences marked. An editor ticks what to accept, and only then is it published.

Answers grounded in your own data. The assistant on this site is tied to the catalog. With each question it is handed the catalog records — the product you are looking at, or an index of all 545 products from which it pulls the full records it needs — and it is told to quote ratios, times and amounts exactly as recorded and never to invent a code or a number. When the catalog does not cover a question, it says so in plain words and marks what follows as general practice, not catalog data.

Checks after the answer. Every product code in an answer is checked against the catalog before it becomes a link. A code that does not exist is not linked, and the conversation is flagged for review. Answers that had to fall back on general practice are flagged too, which shows us exactly where the catalog has a gap.

On your own hardware when privacy matters. For a business the same pattern can run entirely on equipment you own, so that the documents, the records and the questions never leave the building. That is how we set it up when the data is sensitive: client files, prices, contracts.

03Web integration

Clean data only pays off when people can reach it in two taps, on the device already in their hand. In a workshop that is a phone, held by someone wearing gloves.

Search that answers while you type. Type a code, a name or a job, in Russian or English, and the results appear under the field as you type, grouped into products, systems and categories. On a phone the search opens full screen above the keyboard.

Calculators. Every product with a mixing ratio gets a calculator: enter the size of the batch, read off the grams of each component.

A shopping list. Enter the area and the number of coats. The site works out the lacquer, the hardener and the thinner, adds an allowance for waste (15 % unless you change it), converts the amounts into whole packs and totals them across every line of the list. A list can be shared by a link.

A workbench view. A full-screen view for the mixing bench: large numbers, the running total to watch on the scale as each component is poured in, and a screen that is kept awake while you work.

Two languages. Every page exists in Russian and in English at parallel addresses, and search engines are told which is which.

Tools for the editor. Behind the public pages there is an editor's panel: data-sheet import, per-product history with restore, pack sizes, visit statistics that use no cookies, and a log of the assistant's conversations for quality control.