The problem: disinformation does not come in one format
Electoral disinformation in Brazil is not just a text problem. A false claim can travel as a social media post, a doctored image, a video cut out of context or a cloned voice in an audio message. Any tool that only reads text leaves most of that surface uncovered.
That was the starting point for GuaIA, an AI platform created by the Regional Electoral Court of Goiás (TRE-GO) together with the Institute of Informatics at the Federal University of Goiás (UFG). Its goal is to help the Electoral Justice detect and classify disinformation about the electoral process, including deepfakes. I worked on it as a software engineer and co-architect.
What GuaIA does
GuaIA analyzes four kinds of content (text, images, video and audio) and returns an analysis in under four seconds. According to the project page at UFG, the work follows a five-stage flow: extraction, classification, forensic analysis, interpretation and intervention.
It was launched by TRE-GO in August 2024 as a pilot. As reported at the launch, use during the pilot was restricted to the offices of electoral judges and to the state Public Ministry, with wider access planned once the system matured.
My part: services, data and coordination
I shared the architectural decisions with the rest of the team and designed and implemented the Python and NestJS services that make the platform work as a single product. I also integrated several databases and coordinated across the project's different technical fronts.
Three constraints shaped most of the engineering:
- Four formats, one product. Each content type needs its own analysis, but the person using the system should see one coherent answer.
- A latency budget. Under four seconds is a product requirement, not a nice-to-have. It is what makes the tool usable inside a real review workflow.
- A human decides. GuaIA is built as a support tool. The final decision on any measure remains with a person, which is the right design for a system that touches public debate.
Recognition
GuaIA won 1st place in the Innovative Judicial Technology category (Innovative Ideas) at FastLabs 2025, part of the CNJ's Judiciary Innovation Prize, as published by the CNJ. The system has since been adopted, and other regional electoral courts have joined the project, according to its project page.
What I would tell another engineer
If you are building something similar, my advice is to decide early what the product promises (which formats, how fast, with what human oversight) and let those promises drive the architecture. The models will change. The contract with the people who rely on the system should not.
I go deeper into the design questions behind a system like this in the next article of this series.
