Treinei um algoritmo para seguir 80 bilhões de dinheiro público | Redes Neurais Ep. 1: Grafos

Summary

Maria Alice Maia presents the first episode of a three-part technical series about tracing Brazilian public money through open government data. The project combines dozens of datasets into a graph of more than 1.35 million entities and millions of connections. Its purpose is to surface unusual patterns for human review—not to prove wrongdoing or accuse anyone.

How the system works

• Connect data sources: The project links information published across separate databases and in different formats, including election and campaign records, parliamentary amendments, public contracts and payments, company information, sanctions, financial-sector data, and public-servant records.
• Use a graph to follow relationships: Rather than repeatedly joining tables, a graph database represents entities—such as companies, payments, amendments, and campaign donations—as nodes, and their relationships as edges. This makes it easier to trace a money trail across multiple steps.
• Run the data locally: The speaker says the data is stored in encrypted Parquet files on her computer and read with DuckDB. A graph database stores the connections. This approach is intended to handle large datasets while keeping sensitive information off the cloud.
• Check the chain of funds: A simplified path is tax revenue → government budget → amendment → recipient or executor → supplier. The system looks for connected patterns, including possible feedback loops—for example, public money reaching a supplier that later donates to a campaign linked to the amendment’s author. These patterns are treated as signals to examine, not as proof.
• Filter cases for review: Deterministic rules first identify candidate patterns; machine-learning scoring helps prioritize them. The speaker describes a funnel that starts with more than 130,000 pairs, identifies roughly 11,000 with some signal, scores around 2,000 executors, and produces 80 case files for human review. She says she reviews the files before deciding whether to submit a representation to the Public Prosecutor’s Office.
• Preserve traceability: Each graph connection records information such as its source, the rule that created it, and the analysis round. The speaker also says she compared the mapped funds with official totals and found no difference in the amount represented by the map.

Reported observations and cautions

The video illustrates possible signals, such as a company receiving a large first payment shortly after it was created or a financial cycle linking a supplier back to a campaign. It also describes analyzing graph paths and groups of entities that move together, rather than judging companies or transactions in isolation.

The speaker reports that, among amendment authors included in the analysis, unusual connections appeared for 15% of those elected or re-elected, 12% of those not elected, and 2% of those who did not run. She presents this as an observation—not evidence of guilt or a measure of honesty. The names are withheld.

She says some reviewed cases were submitted to prosecutors and that few of the submissions had been archived at the time of recording. The system itself does not determine whether a crime occurred. Its results may contain errors, including mistaken identity matches, and the Public Prosecutor’s Office decides whether to investigate.

Series roadmap

1. This episode: Graph databases, data integration, and tracing relationships.
2. Next episode: How to train and score a neural network with few confirmed examples and no reliable set of “definitely clean” cases.
3. Third episode: How to measure errors and assess the tool’s limitations.

The presenter says the underlying method could also be adapted to other areas, such as health or public procurement, wherever funds leave a trace. She expects to update the system monthly.

Main speaker and sources

• Main speaker: Maria Alice Maia, data scientist, researcher in AI governance, and founder of Micah 6 AI and Estúdio 68.
• Data sources named in the video: Brazilian electoral records (TSE), amendment and transfer records, public contracts and purchases, company and financial-sector records, sanctions, and public-servant databases.