Name: AERTY PINTO DOS SANTOS
Publication date: 09/03/2026
Examining board:
Name |
Role |
|---|---|
| ALBERTO FERREIRA DE SOUZA | Examinador Interno |
| CRISTINE GRIFFO | Coorientador |
| ELIAS SILVA DE OLIVEIRA | Presidente |
| PRISCILA MACHADO VIEIRA LIMA | Examinador Externo |
Summary: This dissertation proposes the enhancement of the rAVA architecture — originally a cog-
nitive textual reader — through the incorporation of advanced artificial intelligence tech-
nologies. The system integrates different forms of reasoning through specialized modules:
Logical Reasoning, Statistics Reasoning, Arithmetic Reasoning, and Look-Up Reasoning,
all coordinated by an adaptive orchestrator called the Reasoning Dispatcher. The new
architecture incorporates Large Language Models, vector-based databases, autonomous
agents, and semantic retrieval mechanisms, while preserving the modularity and scalability
of the original design. Furthermore, the system is now capable of processing multiple
input modalities — such as text, voice, and images — which significantly expands its
scope of operation. This enables deeper data understanding and enhanced decision-support
capabilities across different contexts.
Among the developed applications, the following stand out: (i) natural language queries on
the content of judicial decisions — such as identifying defendants, sentences, aggravating
factors, and legal reasoning; (ii) automatic extraction of relevant information from images
of civil identification documents and procedural records, through optical character recog-
nition and semantic vector search; (iii) voice queries, with automatic speech recognition
and extraction of contextualized legal information; and (iv) symbolic reconstruction of
musical scores from digitized images, using neural networks and classical computer vision
techniques.
Another aspect explored in this dissertation is the use of the system as a predictive tool
in the analysis of judicial decisions, applying Bayesian statistical techniques to estimate
case outcomes and conduct quantitative analyses, such as comparisons between criminal
categories, clustering of judicial decision styles, and detection of semantic patterns in
rulings. Additionally, the use of the system is proposed for identifying judicial biases
through objective metrics and linguistic analyses.
Finally, a proposal is presented for an accessible multimodal interface — also aimed
at ordinary citizens — that allows simplified access to legal information through voice
commands and non-technical language. In this way, the system positions itself as an
innovative and ethical solution for expanding access to justice and modernizing the
Brazilian Judiciary.
