Claude and Ollama helped an Ozon seller process certificates in a day and restore 1,000 SKUs

Ozon began checking certificates for knives before new requirements take effect, and for one seller this ended in a lockdown of almost the entire assortment. Instead of months of manual document uploads, he assembled a working pipeline with the help of Claude and Ollama that processed hundreds of files in a day and helped return more than 1000 SKUs to sale.

Why everything stopped

The author of the case has been selling knives on Ozon for more than six years and describes quality certificates as a constant pain point. The platform regularly changed requirements: in some cases documents were mandatory, in others they were needed only on request, and now Ozon has decided to prepare in advance for new rules that take effect on October 1, 2026. In one day, the marketplace blocked about 1000 SKUs due to missing confirmations.

Previously, the problem was solved manually: go to the certificates section in Ozon Seller, upload a file, then attach it to products one by one. For a few cards this is unpleasant but manageable. For a thousand SKUs, such a scenario turns into months of monotonous work.

The situation was aggravated by the fact that documents were not stored in PDF but as JPEG images on Google Drive, so it was impossible to quickly extract certificate numbers, dates and SKUs from them using ordinary tools.

How they built the pipeline

To avoid burning cloud model limits on hundreds of images, the author moved recognition to a local model qwen3-vl:8b via Ollama. Claude acted as an engineering navigator in this scheme: helping design the architecture, writing scripts, explaining launch steps, and suggesting how to connect OCR with the Ozon Seller API. The key simplification came after checking the documentation: it turned out that images don't need to be published via links or hosted separately.

"The main discovery: the

Ozon API accepts files directly via multipart/form-data".

As a result, the working scenario looked like this:

• a local folder with certificates and attachments was broken down into file pairs;
• Ollama recognized certificate numbers, dates and SKUs from images;
• the script uploaded the files themselves to Ozon via the certificate creation method;
• a separate request got the correspondence of offer_id and product_id;
• the final step linked certificates to goods and saved a CSV report.

Where they had to simplify

The system didn't work the first time. In the process, it turned out that some certificates and attachments were linked by filename, while others were only by the number inside the document. The filenames contained typos, some certificates had SKUs in the attachment, others had them directly on the main sheet.

All of this broke the simple scenario and forced the logic to be more complex, which meant spending more time on the run and more resources from the local model. After several iterations, the author, together with Claude, abandoned overly complex automation and simplified the task. Instead of precise linking of each product within the certificate, it was decided to group products by brand and pass one certificate to Ozon along with attachments, while leaving part of the verification to the marketplace.

Such a compromise immediately accelerated the pipeline: according to Claude's estimate, the recognition time for 225 file pairs and several multi-page documents was reduced from about five or more hours to two or three. The final combat run confirmed that the scheme works not only in testing. Recognizing 231 certificates took 3 hours 23 minutes at an average rate of 52 seconds per document.

Loading to Ozon took another 34 minutes. Only 3 certificates could not be recognized, which is less than 1.3% of the total volume.

If moderation passes successfully, the seller returns more than 1000 items to sale and closes the issue that would have taken months before.

What this means

This case demonstrates well the practical role of AI agents in routine operations: Claude here did not replace the local model, but designed the process, helped fix errors, and brought a user without development skills to a working result. For sellers and small teams, this is a signal that a combination of local vision recognition and API automation is already capable of solving not demonstration, but quite operational tasks.