Training

Training should be conducted after new product registrations or approvals of large amounts of new images. In standard use, the indication is between 2 to 4 weeks.

It is also appropriate to conduct training after a significant number of image additions or corrections.

Good results depend on following good practices.

Metrics

When we receive negative feedback from field operators about poorly recognizing annotations automatically for a given product, and the accuracy metrics within the dashboards were good in recent weeks, it is a sign of not using the annotations in the field. For this reason, we can use the backoffice tool to select and edit the given reports.

After training, it is possible to reuse images from previous reports on mobile, for example by choosing to use images saved in the gallery to see if there has been improvement in cases with negative results.

Rollout

It is important to note that when we start using the automatic recognition system or add a wave of new products, the release to all "Field Operators" should only be authorized after at least one cycle of

continuous improvement in a targeted way, following good practices.

At this initial time a low accuracy is expected due to the need for corrections when editing reports. The results should improve after each Artificial Intelligence training run, and it is recommended to get an

accuracy of 90% or better for most products before releasing the use to all field operators.