Gimba adopts generative AI on AWS to enhance its product catalog.

Gimba CTO (Chief Technology Officer), Daniel Arruda, explains that catalog updates were performed manually until now, for all of the circa 300 new products included monthly. “Each manufacturer sends us their product data in different formats and we have several collaborators who read and standardize this data, creating product descriptions in our catalog, especially the online one,” he explains. João Ricardo Miliozzi David, marketing analyst at Gimba, also adds that “there is a concern with how the product is described as to demonstrate the Gimba personality, inform the consumer about the product’s main features in an exclusive way, and be more attractive to search algorithms, simultaneously.”

According to Arruda, since other Generative AI tools arrived to the market, the team responsible for updating the catalog began experimenting. “We used them in a non-methodological way to produce some texts, which were then revised and improved by the team. This process was manual. Questions were not standardized, nor was the use of the answers,” he says.

With an increase in productivity and product description quality in mind, the Arruda team saw in AWS the opportunity to use Generative Artificial Intelligence (AI) in a more efficient and scalable way. What's more, AWS would be able to take the next step, customizing the AI to Gimba’s so-called “personality”. “We deeply trusted AWS as a strategic partner in our first joint project,” he says.

After presenting the project to AWS, Gimba was invited to develop a platform adapted to its needs and bring the catalog solution to the AWS cloud. With support from Flexa Cloud, the first prototype based on Amazon SageMaker was built for training and adjustments to the LLM model. A sample of 900 products with optimal descriptions was used to train the model so it would understand what is expected. “We tested two or three new products during the learning process and the results were impressive,” says Arruda.

“We knew we were on the right track with this first success” says Deivid Bitti, CEO of Flexa Cloud. “When we received access to Amazon Bedrock, we began the switchover to the AWS managed service.” The chosen model, Claude-2, was critical to the platform's success due to its large context window (of up to 100,000 tokens), allowing us to use advanced prompt engineering techniques that excluded the need to train or manually fine-tune the model. As a result, we reduced the cost of the solution by more than 50%.

The entire development process with AWS was based on the principle of customizing the use of generative AI to Gimba’s needs, specifically those of the registration team. We built an online platform with a single user-friendly interface that automates the application of Amazon Bedrock APIs, and prompt creation on the backend.

With the success of the platform for new products, the registration team expects to pass all the products currently in the catalog through this tool and to reformulate the existing registrations. “By doing this for the new products, we have already freed up the registration team for other tasks, because description takes the most time. This way, we were able to raise the quality standard,” says Juliana.