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

Ozon started checking certificates for knives ahead of new requirements taking effect, and for one seller it ended with nearly the whole catalog blocked. Instead of months of manual document uploads, the seller used Claude and Ollama to build a working pipeline that processed hundreds of files in a day and helped put more than 1,000 SKUs back on sale.

Why everything stopped

The author of the case has sold knives on Ozon for more than six years and describes quality certificates as a constant headache. The platform regularly changed its requirements: in some categories documents were mandatory, in others only on request, and now Ozon decided to prepare in advance for new rules that take effect on 1 October 2026. In a single day the marketplace blocked about 1,000 SKUs for lacking the required confirmations.

The problem used to be solved by hand: open the certificates section in Ozon Seller, upload a file, then link products to it one by one. For a few listings that is annoying but tolerable. For a thousand SKUs, it turns into months of monotonous work. The situation was made worse by the fact that the documents were stored not as PDFs but as JPEG images on Google Drive, so ordinary tools could not quickly extract certificate numbers, dates and article numbers from them.

How the pipeline was built

To avoid burning through a cloud model's limits on hundreds of images, the author moved recognition to the local qwen3-vl:8b model via Ollama. In this setup Claude acted as a navigator-engineer: it helped design the architecture, wrote scripts, explained how to run each step and suggested how to connect OCR to the Ozon Seller API. The key simplification came after checking the documentation: it turned out the images did not need to be published via links or hosted separately.

"The main discovery: the

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

The working workflow ended up like this:

• a local folder of certificates and appendices was split into pairs of files;
• Ollama recognized the certificate number, dates and article numbers from the images;
• a script uploaded the files themselves to Ozon via the certificate-creation method;
• a separate request fetched the mapping between offer_id and product_id;
• the final step linked certificates to products and saved a CSV report.

Where it had to be simplified

The system did not work on the first try. Along the way it became clear that some certificates and appendices were linked by file name, while others were linked only by the number inside the document. File names contained typos; for some certificates the article numbers were in the appendix, for others right on the main sheet. All this broke the simple workflow and forced the author to add processing logic, which meant more run time and more local-model resources.

After several iterations, the author and Claude abandoned overly complex automation and simplified the task. Instead of precisely linking each product within a certificate, they decided to group products by brand and submit one certificate to Ozon together with its appendices, leaving part of the verification to the marketplace. This compromise sped up the pipeline right away: by Claude's estimate, recognition time for 225 file pairs and several multi-page documents fell from a little over five hours to two to three.

The final production run confirmed that the approach works outside of testing. Recognizing 231 certificates took 3 hours 23 minutes, averaging 52 seconds per document. Uploading to Ozon took another 34 minutes. Only 3 certificates could not be recognized, less than 1.3% of the total. If moderation passes, the seller puts more than 1,000 products back on sale and closes a task that would previously have dragged on for months.

What it means

The case shows the practical role of AI agents in routine operations: Claude did not replace the local model here but designed the process, helped fix errors and took a user without development skills to a working result. For sellers and small teams, it is a signal that combining local vision recognition with API automation can already handle operational tasks, not just demos.