My squad is a cross-functional team responsible for developing, serving, maintaining, and improving machine learning models and checkpoint services for fraud prevention in the Global market. We screen all Card Present and banking transactions at Sumup around the world (BR, US, UK, EU, LATAM, AUST). As a Data Science and Engineering Lead, my primary objective is to enhance team performance, enabling them to deliver higher-quality results.

My squad was responsible for developing, serving, maintaining, and improving machine learning models for fraud prevention in the Brazilian market. As a Data Science Lead, my primary objective was to enhance team performance, enabling them to deliver higher-quality results.Key Projects:- Payment Link Fraud Detection: Implemented machine learning algorithms to identify and prevent fraudulent transactions in payment links. We developed, deployed, ran A/B testing, and created the monitoring and retraining process for the model. This initiative increased company revenue by nearly 20%.- Card Present Fraud Detection: Developed models by exploring complex statistical features, deployed them, and conducted online validation. The evaluation showed that the model reduced fraud by 50%.- Team Performance Improvement: We developed an entirely new machine learning pipeline that enabled the team to deliver better and more. This included centralizing common code, building predefined pipelines as a first plug-and-play model trial, and developing step-by-step guides to deploy a model.I also played a key role in structuring the squad, including organizing team members, developing Y-career paths, career tracking, and project prioritization and definition.

I was responsible for enhancing the performance of the data science team through the creation of processes, culture, and infrastructure. I reduced the time required to develop and deploy a model from one year to one month. In this role, I structured the entire management of machine learning projects/models by creating Data Science Project templates, defining repository policies, implementing CI/CD for automation, and developing a new model training pipeline. Additionally, I implemented processes for monitoring and validating models in production to ensure quality and robustness.

At Hurb, I am responsible for identifying and refining AI demands, modeling, deploying, and maintaining machine learning products. Throughout this experience, I participated in challenging projects:- Search customization:Multi-modal (text and image) model based on Dual Encoder to personalize our search results.- Search Filter improvement:Parts of Speech recognition and LLM in semantic search approach to matching hotel to filter topic.- Image intelligence:Bidirectional Encoder representation from Image Transformers to extract labels from images (24 Million images).- Causal Inference:Synthetic contra factual strategy to measure pricing strategy effect, using Google Causal Inference lib.- Activity recommendation system:Multi-modal (text and image) model based on Dual Encoder to deduplicate and suggest better activity tickets to our clients.- Next trip recommendation:LLM to vectorize Wiki data and recommend the next trip, based on destination similarity.And I helped in the structuring and development of the team:- Software engineering reading group:Reading group between areas, bringing a clear improvement in the overall quality of the team's codes- Data product discussion group:Internal and external company guests with product experience share their know-how, bringing a product bias to our projects.- Optimization of internal processes:Standardization of PRs reviews, relationship and project management, and project template

In the LMDC, I led the development of advanced machine learning and deep learning models for a research project with the biggest Brazilian oil company. My goal was to assist the company in quality control of seismic acquisitions in oil fields, developing:- Data analysis:Dealing with unusual model format (SEGY file) e building visualizations to signalize data issues.- Valid or invalid seismic acquisition:Analyzing modern classification (XGBosst, CatBoost, etc) and traditional (Rand, Forest, MLP, etc) to classify seismic acquisition into invalid or not.- Time series analysis:Dynamic time warping (DTW) to cluster seismic signals.- Seismic Signature Translation:Machine translation (Encoder-Decoder) strategy adaptation to translate seismic signals into deep water signals.

At Visagio, I participated in management projects linked to technology, covering different sectors such as retail, real estate development, and bike sharing. I developed solutions, working like:- Pilot-Route-Bus allocation:Constraint optimization model to define which pilot would drive which bus on which route.- Credit classification:Machine learning techniques to predict churn in real estate development.