The route of one line

A single item from a delivery note, seven stations long. The left side shows what happens at each point; the right side, what the data looks like afterwards. Newly added content is copper-coloured, discarded content struck through.

Station 1 of 7 –

Brand, numbers, sizes, quantities and prices in this demonstration are made up. “Steinlauf” does not exist, the number starts with a sample prefix, and the amounts are placeholders. The process, on the other hand, is the real one.

Three steps, deliberately separated

One costs money, one changes the shop, one does neither. Combined in a single button, they could no longer be undone individually.

Step 1

Read in incoming goods

The delivery list is checked line by line against the existing stock: new, partly present or long since in the shop.

costs nothingchanges nothing

Step 2

Research

Only for items that are genuinely new, and only after selection. The estimated amount is shown next to the button before anyone presses it.

costs moneychanges nothing

Step 3

Create

In batches, throttled, with a safety prompt. Every product is created inactive and with zero stock.

costs nothingchanges the shop

What is not shown here. The tool works on the real stock of a real shop. This page therefore contains no product numbers, prices, supplier data or images from it. The demonstration above runs on invented data.

The interface

Everything is operated in the browser, on the computer in the shop. The process is laid out as a row at the top: scan, incoming goods, research, upload – it is always visible where one stands and what the next step costs.

Interface of the tool: header with the four steps Scannen, Wareneingang, Recherche and Hochladen, below it an input field for the product number and a short guide in four columns. — Captured without a running service – which is why “Dienst antwortet nicht” (service not responding) appears at the top and the lists are empty. That is exactly the point here: what can be seen is the design, not a single record from the shop.

Five pages, clearly separated

Page	Purpose	Affects
Create	The main route: incoming goods, research, review list, upload	reads the live shop, writes to the test system
Descriptions	Review existing texts and plan a rewrite	reads the live shop, writes only locally
Product numbers	Review numbers, assess them and note down the rule behind them	reads the live shop, writes only locally
Access	The brands’ B2B portals: access, conditions, copyright notice	local only
Manual	Operation, step by step	–

The table has three columns – on a phone it can be swiped sideways.

Target system indicator. The tool can write to two systems. The header therefore shows at all times which one is currently being worked on. The indication comes from the connection itself – not from a label that someone might forget to switch over.

Eleven stations

From the delivery line to the post-check in the shop. The right-hand column shows whether a station costs money and whether it changes the shop.

Station	What happens	Consequences
00 · Catalogue copy	The stock is fetched read-only as a local copy.	read-only
01 · Delivery list	Old character encoding, messages before the header row, corrupted identifiers – detected rather than guessed.	changes nothing
02 · Matching	Multi-stage duplicate check. Result: new, partly present or present.	changes nothing
03 · Categorising	Categories, size axis and properties from the most similar existing products. Without a language model.	changes nothing
04 · Research	Language model with web search, only for selected products. Limits per product and per day stop the run.	costs money
05 · Proposal	Title, description, product number and mandatory details in the house style.	changes nothing
06 · Review	Traffic light per field, concerns at three levels.	changes nothing
07 · Approval	A person sets the price and approves. Without a price, nothing is created.	Lock
08 · Create	One call creates the parent product, sizes, categories and properties. Inactive, zero stock.	changes the shop
09 · Post-check	Every product is read afresh and compared against the proposal.	read-only
10 · Rollback	Files, drafts and jobs can be discarded individually.	reversible

Where a product can fail

On reading in. Corrupted product identifier, wrong character encoding, messages before the header row, size lines without grouping.
Against the stock. Duplicates at several stages, authenticity of the identifier via the allocation range, attachment target for missing sizes, age of the catalogue copy.
At research. Consistency between product type and image, weight against size range, source requirement, hard cost caps.
At the image. Origin and right of use per brand, format, background removal with metrics, deep check, follow-up check.
Before creation. Traffic light per field with a plain-language reason: stop, warning or note.

A matching check digit proves nothing

Anyone who assigns numbers themselves usually calculates the check digit correctly. What is checked, therefore, is the allocation range: ranges reserved for in-house use must not enter the shop as an official identifier.

Why categorising without AI is better

The question “where does a softshell jacket belong?” has long been answered in the stock – hundreds of times, by people who know the shop. Having a language model choose from several hundred categories costs money, takes time, and at best justifies itself with “because it fits”. Matching against the most similar existing products costs nothing and justifies itself: because 214 other softshell jackets are there.

Images

Three separate routes, distinguished by how strongly they intervene.

Route	What it does	When
Standard	Cropping, squaring, white background, uniform edge margin. Purely computational, without AI.	on every image
Deep check	Five methods search for markings and measure image defects. Changes nothing.	in the review screen
Image editing	Manual re-cutting when the automation reaches its limit.	on request

What the deep check looks for – and why it removes nothing on its own

It looks for copyright notices in the metadata fields, for lettering, for faint overlays, for patterns in the least significant bit and for repeating tiles. All algorithmic, without any outside call. Each of the five searches also states what it cannot do.

Every find gets a frame, a number and – for invisible finds – an evidence image. Then the tool asks, and without an answer nothing happens. A find classified as third-party blocks the image. Only what a person has confirmed as their own mark is removed. A third-party photo does not become one’s own through editing, and a removed copyright notice makes the situation worse.

Measurement is done on a test bench with self-generated test images. Half of the cases check whether the search can stay silent – a detector that goes off on every image is as useless as one that never goes off. This uncovered five real bugs, including a search that mistook the ventilation slots of a helmet for a pattern.

Background removal without AI

Instead of a fixed brightness threshold, a model of the background is fitted to the edge pixels – a slightly tilted plane with fall-off towards the centre. This covers lighting gradients and vignetting. The tolerance comes from the spread of these edge pixels: a clean light tent gets a tighter one than a wooden table. White on white without contrast is reported as unsuitable rather than guessed.

The rule sets

Read off the existing stock, not invented. Their content is business know-how and is not given here – what is described is what they decide.

Rule set	What it decides
Number structure	What the product number of a new model has to look like – and when, by exception, it differs.
Duplicate check	Whether the product already exists. Several stages, from “proof” to “hint”.
Categorisation	Categories, size axis and properties – from the most similar existing products.
Number authenticity	Whether a number is an assigned identifier or invented in-house.
House style	Structure of title and description, what belongs in them and what never does.
Image source per brand	Which dealer access the material comes from and which copyright notice is carried along.
Cost limits	How much a research run may cost per product and per day.

What else it can do

Much of this came about during building, because a closer look revealed that something was missing.

Add sizes instead of creating duplicates. The most common case when reordering. Title, images, categories and price remain untouched.
Read broken export files. The reader falls back on another field instead of passing on nonsense.
Remove backgrounds from in-house photos. Without AI, using a model of the background instead of a fixed threshold.
Review existing titles and texts. A large part can be fixed without AI and costs nothing. Only the rest goes into the cost estimate.
Manage dealer accesses. Portal, conditions and copyright notice per brand.
Total up stock levels. One run sets the parent product to the sum of its sizes.
Add mandatory details. The manufacturer information is tied to the brand. Existing data is never overwritten.
Undo everything. Whatever is already live in the shop deliberately stays – with a notice rather than silently.
Compete against manual upkeep. The proposal next to a manually maintained product. Where they differ, either the rule is wrong or the stock is.

Status

Built in August 2026, component by component. The core works.

Surveying the stock	Done
Reading, mirroring, checking	Done
Categorising and proposal	Done
Write path and rollback	Done, on the test system
Images from the dealer portals	Done
Image check and manual editing	Done
Plausibility and title review	In progress
From the test system to the live shop	Open – it remains to be clarified which system holds which product data

The tool replaces no one; it takes over the typing. The price is never set automatically. It runs on one computer, without a network service, and accesses the shop server read-only. Image material comes from the manufacturers’ B2B portals or from in-house photography – an edited image is not free of rights. The tool is shown with the consent of Muskelkater Sport Köln and without data from ongoing operations.

Who this is built for

Starting point: Incoming goods arrive as a delivery list. Creating products from it by hand takes a long time, and the longer it takes, the more inconsistent the data in the shop becomes.
Target group: Shopware operators with regular deliveries and a catalogue that follows its own house rules: number structure, size axes, categories, spellings.
Status: In development. Core, interface and test benches are running; the write path has been run through on a test system. The tool is not yet in live operation.
Availability: On request, as a custom implementation. The tailoring depends on the catalogue: the rules come from the existing stock, not from a preset.
Next step: Running through the process above once shows whether it fits a shop’s incoming goods. If it does, a message is welcome.

Run through the process Get in touch

From delivery note
to finished product.