Jose Luis Moreira Arruda

A machine learning system to monitor beer quality during and after production. We perform the ETL process of thousands of industrial sensor signals using SQL. Also, we deploy machine learning models to predict the quality of the beer in real time and identify correlated root causes for production problems or improvements.

Our system also performs simulations, giving real-time visibility to the brewers and a way to act in real time during the process and improve product quality. It was deployed in the Azure/Databricks environment. The project was deployed in Brazilian breweries and mainly impacted the improvement of beer quality and waste reduction, saving over $1 million per year.

Developed a new approach to extract image information from products for a big fashion marketplace. Leveraging the amount of data and self-supervised techniques, we could reduce human bias and capture rich features as texture improved other downstream systems.

My colleague and I competed in one of the largest international competitions for forecasting, the M5 competition, in 2020. We won a gold medal for 7th place among 900+ teams worldwide.

That year, the competition task was to forecasts a daily demand distribution in quantiles for each Walmart product and three Walmart stores. This was a combination of over 40,000 time series.

Since we wanted to make our solution more general and practical and turn it into a product, we developed a unique end-to-end deep learning model based on recent advances of transformers for time-series forecasting instead of multiple ensembles or other nonsuitable models for production approaches.

A ranking system for one of the world's largest luxury marketplaces. While developing this ranking system, I considered multi-modal data, multiple business requirements, and relevance to each user to improve conversion, profitability, and other target metrics.