Brenda Oliveira Ramires

Demand forecasting models that helped a company expand its presence in Brazil by opening additional stores. We used aggregate information about competitors that the client purchased from a consulting firm, the client's own data, and the data the client collected about Brazil to model the demand per region, and how much of that demand was satisfied by the competitors. The result was a map highlighting areas with great potential for new stores.

A data-driven CRM solution for a retailer in Brazil. The company wanted to use the data from previous purchases to identify types of customers and create more personalized discounts. The solution used basket analysis to identify products that were often bought together and analysis of past customer behavior to identify which types of discounts had worked better and in which phase of the relationship with the brand the customer was; for example, recently started the relationship, coming back after some time of not shopping, or highly loyal.

In the end, the solution identified the best products to apply a discount to and the clients that needed to receive the discount in order to achieve a goal from the business side; for example, make the client loyal to the brand, retain a casual buyer, or increase average ticket.

A pipeline that collected, cleaned, and organized data to be used in different projects by all the other data scientists on the team. Using crawlers and public databases, we collected data to be used in our projects. Each team identified databases from which to extract data and processed the datasets in their own way. As a result, teams often processed databases that had already been processed by other teams.

To save time for our data scientists, I established a team to centralize data collection and processing. We automated the execution of crawlers after creating a standard for how a crawler should function and the output it should generate. We saved this first output and the form we created with important features that we made available to everyone. We kept the first output because we could always create more features from the original dataset. In the end, we all had one place to go to look for data, and we didn't have to waste time processing the same data again.