Borjomi of the first in Russia was implemented by a neuronet for work with retail

The SmartMerch company announced on November 27, 2018 TAdviser end of the main project stage in IDS Borjomi Russia on automation of trade audit based on neural networks and machine learning technology. A project objective is quality improvement of collecting of marketing information. According to representatives of SmartMerch, it is the first in the Russian Federation commercially successful example of use of AI for automation of similar tasks. As of November 27 the accuracy of recognition of products of Borjomi in outlets reaches 93%.

Borjomi Russia — one of the most active players in the Russian market of butylated mineral water. That buyers were happy and quickly found our products in outlets, we should provide completeness of the range and observance of standards of the calculation. Merchandisers are responsible for control of shelf space in retail. Before SmartMerch implementation they visually estimated the calculation and manually fixed information. This, collected thus, often are incomplete and inaccurate because of the large volume of works and influence of a human factor. Therefore IDS Borjomi Russia made the decision to automate trade audit using a sensing technology of images. We held the tender and after successful "pilot" with his winner — the SmartMerch project — started scaling all over the country — Artur Markaryan, IT Business Partner IDS Borjomi Russia told.

A system which cornerstone the neural network is processes pictures, will recognize the goods represented on it and prepares the reporting on quality of the calculation, observance of plano-grams, completeness of the range, existence of out-of-stock and so on, Maxim Arkhipenkov, the creator and the SmartMerch project manager explained. At Borjomi the business model on application of such systems was ready: the company looked for the supplier capable to provide correct data in terms of quality. Therefore within "pilot" a system was tested in "the most fighting mode" — under loading, with a large number of SKU and high detailing.

According to Arkhipenkov, within the project in Borjomi the stage of predata analysis and basic training of a neuronet at test photos took 15 days. Further iterations of after-training for the purpose of further quality improvement were carried out. As a result during "pilot" it was succeeded to reach the high accuracy of recognition of the image of 95-97% at time of processing of one picture — 5 seconds.

Already at the initial stage of a pilot project it became clear that SmartMerch yields much with higher quality results, than a human eye. However results even exceeded our expectations. We see that a system allows us to obtain field data of very high quality and in short terms for operational acceptance of management decisions and we can recommend use of SmartMerch service to the companies interested in high-quality carrying out and the analysis of trade audits — Artur Markaryan added.