Thiago Meirelles Ventura

This study investigates the use of the BERT model to support the automated classification of public documents according to whether they contain personal or sensitive data, in compliance with the Brazilian General Data Protection Law (LGPD). Documents from the Mato Grosso State Official Press (IOMAT) were used, comprising 1,050 documents labelled as sensitive and 450 labelled as non-sensitive. In this version, only a document-level binary classifier (sensitive vs. non-sensitive) was implemented and evaluated; Named Entity Recognition (NER), Explainable AI (XAI), and an auditing interface are described as part of a broader system planned as future work. The model — a fine-tuned BERTimbau classifier — was trained after removing explicit date expressions from the texts, with a fixed seed and deterministic settings for reproducibility. On the validation set, it reached 75.00% accuracy and an 83.66% positive-class F1-score, with a macroF1 of 65.23%. Compared with a majority-class baseline, the main improvement was observed in macro-F1, indicating that the model learned useful patterns beyond simply assigning all documents to the sensitive class. An error analysis based on the confusion matrix shows that false negatives are the legally costlier errors in a compliance setting, as they may correspond to sensitive documents being treated as non-sensitive. The results should be interpreted as preliminary and corpus-dependent, since date removal reduces but does not eliminate temporal and topical composition effects. Even so, the findings indicate the feasibility of the approach as an initial, human-in-the-loop screening stage for ethical data governance in the public sector.

The adaptation of foundation language models for legal document generation has become a relevant problem in Legal Natural Language Processing, particularly because generated texts must preserve factuality, normative coherence, traceability and control over sensitive information. This systematic literature review maps, classifies and critically analyzes primary studies on domain adaptation techniques for foundation language models applied to legal and normative-argumentative document generation. It also examines the implications of the resulting evidence for analogous regulated institutional contexts that share requirements related to reliability, privacy, auditability and institutional control. The review followed guidelines for systematic reviews in Software Engineering and the applicable items of PRISMA 2020. Searches were conducted in Scopus, Web of Science, IEEE Xplore and ACM Digital Library, covering studies published from January 2021 to December 2025 and complemented by backward snowballing. From 200 initial records, 32 primary studies were included. The corpus is predominantly concentrated in the legal domain, with only one study directly addressing document production in Public Security. The results show a recurrent use of hybrid adaptation strategies involving parameter-efficient fine-tuning, retrieval-augmented generation and structured inference, although the literature does not establish a universally superior combination. Data engineering is marked by a trade-off between scale and control, with frequent use of semi-synthetic datasets. Traditional Natural Language Processing metrics remain useful as initial indicators, but are insufficient to assess factuality, normative validity and argumentative consistency. Ethical safeguards, privacy controls and bias mitigation metrics remain underdeveloped. The findings indicate that the institutional adoption of foundation language models for legal document generation should be treated as a problem of architecture, evaluation and governance, rather than only textual performance. These conclusions are directly grounded in the Legal NLP literature, while their transfer to analogous regulated institutional contexts should be understood as a functionally motivated implication requiring domain-specific empirical validation.

The upkeep of roadside vegetation is essential for ensuring the safety of both motorists and pedestrians. However, identifying the necessity for such maintenance is frequently a time-consuming and costly process, with a high potential for errors in annotation. This work therefore proposes the development of a solution capable of estimating the height of vegetation along roadsides in an automated manner. A machine learning model was developed and evaluated using a dataset of manually annotated data. The model employs a convolutional neural network architecture, adapted for the task of classifying vegetation heights. The results demonstrate that the model is able to detect the different vegetation height classes with low error rates, indicating its potential for automating the decision-making process for mowing vegetation. This study contributes to the advancement of road monitoring techniques, providing greater operational efficiency and reducing costs for road maintenance.

Este artigo propõe um modelo analítico de governança para orientar o desenvolvimento colaborativo de soluções de Inteligência Artificial na segurança pública em parcerias público-acadêmicas. A pesquisa, de natureza qualitativa, fundamenta-se em revisão bibliográfica e análise documental de normas como ISO/IEC 27001, ISO/IEC 42001 e o NIST AI RMF, além da legislação brasileira de inovação e propriedade intelectual. O modelo estrutura-se em uma base de gestão de riscos e quatro pilares: segurança da informação, gestão do ciclo de vida da IA, propriedade intelectual e conformidade legal. A proposta busca alinhar a inovação tecnológica à transparência, à governança responsável e à autonomia tecnológica.

Este trabalho investiga a viabilidade do uso de técnicas de machine learning aplicadas à bioacústica para identificação automatizada de aves ameaçadas no bioma Pantanal. Dados bioacústicos foram coletados da plataforma Xeno-canto e utilizados juntamente com redes neurais convolucionais e mel spectrogram. Os resultados demonstraram desempenho satisfatório, com valores de recall que variaram entre as espécies selecionadas: Anodorhynchus hyacinthinus (0,94), Calidris canutus (0,93), Calidris pusilla (0,80), Crax fasciolata (0,70) e Harpia harpyja (0,89). Concluiu-se que a aplicação de machine learning para detecção de aves ameaçadas é tecnicamente viável, mesmo para espécies em extinção.

This work presents a series of experiments aimed at recommending scientific events within the Brazilian Computer Society (SBC), leveraging semantic embeddings generated by Large Language Models (LLMs) and unsupervised clustering techniques. Using BERTopic for topic modeling and multilingual representations, the experimental approach processes over 12,000 articles from 30 events to align submissions with relevant themes. The best configuration reached an accuracy of 0.91 and demonstrates the potential of LLM-based embeddings to support decision-making in scientific dissemination.

This article details the construction and validation of the GolpeBR dataset, which was created from news articles and Reddit posts. Automated Python methods were used to extract and process the data, which was then annotated using the Deepseek-R1 LLM and the 5W1H methodology. A cybersecurity expert classified and validated the records, distinguishing between banking cybercrimes and unrelated crimes. For the dataset's validation, supervised learning algorithms were applied. It was found that models trained with data structured by the 5W1H methodology demonstrated better accuracy, reaching 0.83 for the Logistic Regression and Random Forest algorithms.

This article aims to develop a model for summarizing police reports in the context of public security, with a focus on execution on limited hardware. The approach combined hybrid distillation (logits and intermediate representations) and supervised fine-tuning with LoRA, using a corpus of 19,286 police reports. The evaluation was conducted using automatic metrics (BERTScoreF1) and qualitative analysis by specialists. The results demonstrate that the distilled model generates clear, coherent, and semantically appropriate summaries, comparable to those of the larger model, with superior computational performance, including in hardware-constrained environments.

Introdução: A descrição automática de imagens por sistemas de inteligência artificial é essencial para a acessibilidade de pessoas com deficiência visual, mas pode reproduzir vieses e comprometer a representatividade. Objetivo: Analisar implicações éticas na geração de descrições, considerando apagamento identitário, viés racial, estereótipos de gênero e invisibilização da deficiência, usando ChatGPT e Gemini. Metodologia: Estudo qualitativo e exploratório com sete imagens contendo marcadores identitários, descritas pelos dois modelos e analisadas por conteúdo temático. Resultados: Algumas descrições, embora coerentes, apresentaram omissões, linguagem genérica e estereótipos, apontando a necessidade de diretrizes para descrições mais representativas e inclusivas.