João Rafael

A real-time credit-card fraud detection system for one of the largest payment processors in the US. This system offered the client an API for scoring transactions and a fraud score with automatic explanations for that score. This was a multi-datacenter and multi-tenant project with hard latency requirements and mandatory high-availability.

In addition to defining the solution's architecture, I coordinated and communicated requirements, progress, and deadlines across the client's technical staff, the company's project managers, and internal product engineering and research teams; ensured high-quality code and documentation, and coached each team member through one-on-one meetings.

A product recommendation system for a large telecom. This product was used to define the upsell marketing strategy for existing customers and would recommend specific upgrades and bundles for each client, such as voice cards, mobile data, and premium TV channels.

I developed this system from scratch. A content-based approach was used to incorporate information from multiple domains, including product usage, billing information, previous customer interactions, and demographics, as well as product characteristics and price points. The recommendations were measured against the previous strategy in A/B tests, and a statistically significant increase in average revenue per user (ARPU) was achieved.

A service that optimizes call centers' outbound operations (e.g., marketing campaigns) by identifying the best time to call each client. The system uses sociodemographic information together with the history of past call attempts to predict the best time to contact each client. Additionally, a global optimization process is used to define the best order in which to call clients to maximize the operation's KPI performance (e.g., client reach or right-party answer rates).

The system communicates with multiple partner companies and accesses several back-end systems and databases to collect the necessary information. A dashboard was created to monitor both the system and the final business metrics.

I led the development of the process from proof-of-concept to production, ensuring correct development processes and code quality by means of unit and integration tests, code linters, CI/CD, and code reviews.

This project drove a 15% increase in client reach for the selected marketing campaigns. It was showcased as a case study for the data science community and presented to an audience of 80+ data scientists, industry players, and C-level executives.

A system to detect and predict degradation in the quality of service of cable, fiber internet, and TV signal delivery for a telecom servicing over two million devices. This system monitors technical low-level metrics such as transmission signal to noise ratio, pathloss and error rates, and other information related to the network stack such as IP availability, network topology, and throughput and latency. It also processes semi-structured information obtained from real-time logs of the management infrastructure and the devices themselves.

The model combines anomaly detection and predictive algorithms to identify which clients are facing or will face network issues. Due to the large amount of data collected, PySpark was used to process the information in a cluster. Specific code was developed in Java for extra optimizations.

Throughout this project, several distinct patterns were discovered in the data and relayed to the company's engineering team to fix. Additionally, a survey was conducted, contacting the clients who were most likely to be facing problems, and 98% confirmed our findings.

A computer vision project that used satellite images and official records of building footprints to detect viable locations for solar panel placement on rooftops. I completed the initial milestone of the project in two weeks, and it outperformed existing individual human labels.

A sales forecast model and data exploration tool for a beverage producer and distributor. The model would forecast sales of hundreds of individual SKUs for the next three months at each point of sale, including coffee shops, groceries, and supermarkets.

I implemented the project, which included discussions with the client data and product teams to understand and clean data issues. Datasets were enriched with external data sources for weather, demographics, and events. The final deliverable included a dashboard where the client could visualize the data geographically and uncover patterns across locations and time spans.

I was responsible for the design, implementation, optimization, and maintenance of the client's data lake solution, which empowered the business's data needs across product, operations, analytics, and AI functions.

The data lake supported both bulk and streaming data ingestion from various sources, including data brokers, product usage, SaaS services, operational logs, and APIs. I ensured data cleanup and transformation processes were in place to meet both operational and analytical requirements effectively.