Experience

I work in the Corporate Security Modelling team at Bradesco. My role involves developing and monitoring machine learning models to detect and prevent fraud.

• Uncovered a major vulnerability involving high-risk accounts, revealing R$ 5.2 million per month circulating through accounts that should have been blocked, leading to strengthened fraud-prevention controls.
• Built a scoring model to detect mule-account behavior at onboarding branches, delivering R$ 3.8M/year in fraud-prevention impact by improving alert prioritization.
• Developed a fraud-profile model for business onboarding that delivered a 30–40% increase in detected fraudulent account-opening attempts, significantly strengthening early-stage fraud prevention.
• Developed a behavioral detection model analyzing 19.6 million active accounts, combining behavioral and demographic data to identify high-risk profiles, ultimately isolating a refined set of highly suspicious accounts.
• Led a large-scale fraud-detection initiative on a PIX transaction network (4.5M nodes, 6.5M edges), uncovering hidden intermediaries and exposing mule networks.

Thesis: Network science and machine learning applied to criminal networks

I conducted research on network and data science applied to corruption and organized crime, focusing on extracting meaningful patterns from criminal activity data. My work aimed to uncover underlying structures, rules, and mechanisms shaping these networks, contributing to a deeper understanding of their dynamics.

Thesis: The dynamics of political corruption networks

I explored the structural evolution of political corruption networks using network science and data science. My research revealed unexpected patterns, including linearity in community structure evolution, trends in the growth of repeat offenders, and a coalescence-like process in network formation. Based on these statistical similarities, we developed a corruption network model capable of replicating empirical findings.

Thesis: Time series analysis via complexity-entropy curves

I introduced a novel parameter, embedding delay, to analyze time series from various sources. By extending the complexity-entropy curve approach, this method proved effective in detecting periodic patterns within noisy signals.