Sérgio Colcher

This article presents the technical specifications resulting from R&D and standardization efforts in the application coding layer of the next generation of the Brazilian Digital Terrestrial Television System (SBTVD), known as TV 3.0 or DTV+. The architecture evolves the Ginga middleware into an application-oriented platform, specifying APIs and definitions that enable personalized experiences, integration with Over-The-Top (OTT) services, new forms of interaction and immersion, enhanced accessibility, and audience data collection with privacy safeguards. The initiative is the result of collaboration among the SBTVD Forum, its associated companies, the Ministry of Communications, the National Education and Research Network (RNP), and Brazilian universities.

Developing interactive hypermedia applications using the Nested Context Language (NCL) presents significant challenges due to its declarative nature and strict syntactic and semantic requirements. This paper introduces a multi-agent tool powered by large language models (LLMs) to assist in the authoring of NCL 3.0 documents. The proposed tool leverages agents specialized in generating, validating, and refining NCL elements, ensuring syntactic correctness and structural alignment. The tool’s usability, workflow efficiency, and perceived user satisfaction were evaluated through a structured observational study. This work contributes to the field of multimedia systems by providing a scalable, efficient, and user-friendly solution for NCL development, paving the way for the broader adoption of declarative languages in digital TV and multimedia applications.

Smart contracts are fundamental to blockchain ecosystems, but remain susceptible to security vulnerabilities that can lead to severe financial losses. Recent advances in agentic AI systems, powered by large language models (LLMs), enable autonomous code analysis and decision-making without explicit task-specific supervision. These systems leverage prompt engineering and zero-shot reasoning to detect vulnerabilities in smart contracts without prior fine-tuning. In this work, we evaluate the effectiveness of agentic LLM-based approaches in identifying vulnerabilities using prompt engineering and zero-shot reasoning across a curated dataset of Solidity smart contracts. Our findings highlight the limitations of current LLMs in automated vulnerability detection, providing insights into their practical applicability for securing decentralized applications. Our best-performing configuration, which integrates zero-shot reasoning with the Tree of Thoughts framework, achieved an F1-score of 73.66%.

The field of narratology has long worked on different ways to classify characters according to their narrative role and importance; meanwhile, studies on character design have observed the existence of unclear patterns that correlate specific visual features and those narrative classifications. Therefore, it becomes worth considering if modern image classification methods can be used to not only achieve satisfactory performance on this task, but also help us learn about the patterns that shape it. The task of classifying characters carries a vast array of practical uses, such as creating recommender systems based on specific visual-narrative correlations, automatic annotation for media preservation, adaptive interactive storytelling based on how a character is perceived, and the creation of education tools on how those patterns change depending on culture. In this work, we construct a dataset customized to assessing both performance and visual–narrative correlations within a specific medium, conduct an anonymous survey to evaluate human performance and tendencies, and compare the results with those of a traditional CNN classifier and a Foundation Model–augmented approach. Both models outperform the survey participants, with the latter achieving 76.93% accuracy, 18.11% higher than human performance, while also providing insights into the visual cues most closely tied to the given narrative roles.

Recent advances in self-supervised learning have significantly improved visual representation learning, creating better alternatives to supervised methods in image understanding tasks. Motivated by the demand for robust features under low-quality image scenarios, we propose adapting the SimCLR framework to operate in the frequency domain using quantized DCT coefficients from JPEG-compressed images. We aim to investigate the integration of the Discrete Cosine Transform (DCT) with self-supervised contrastive learning to enable visual representation learning. To evaluate the learned representations, we apply linear classification and JPEG artifact removal tasks, employing ResNet-50 and Vision Transformer (ViT) encoders. Results indicate that ViT-based embeddings, especially when combined with spatial or frequency-domain inputs, mitigate performance loss degraded by compression. Furthermore, artifact removal experiments with modified UNet architectures show that incorporating ViT embeddings and attention maps can slightly improve reconstruction quality. These findings suggest that frequency-domain self-supervised learning is a promising direction for building robust and transferable visual representations. However, the observed gains remain modest and limited by computational constraints.

The evolution toward TV 3.0 establishes a hybrid broadcast–broadband ecosystem, application-oriented and characterized by new requirements for personalization and interoperability. Although recommender systems are widely explored on OTT platforms, their use in the television context still faces specific challenges, such as limited interaction, shared viewing, and constraints on data collection. This work proposes a conceptual model for a recommendation service based on multiple specialized agents coordinated by a central orchestrator. The proposal aims to provide a minimum viable abstraction, modular, extensible, and aligned with the constraints of the television environment. In addition to detailing the architecture, we present a hypothetical use case that illustrates the end-to-end flow and discuss implications for transparency, explainability, and governance of recommendations. The results are intended to support the adoption of personalization mechanisms in TV 3.0, reconciling technological innovation with requirements for reliability and usability.

This work presents an unsupervised method for tagging banking consumers’ transactions using automatically constructed and expanded topic taxonomies. Initially, we enrich the bank transactions via web scraping to collect relevant descriptions, which are then preprocessed using NLP techniques to generate candidate terms. Topic taxonomies are created using instruction-based fine-tuned LLMs (Large Language Models). To expand existing taxonomies with new terms, we use zero-shot prompting to determine where to add new nodes. The resulting taxonomies are used to assign descriptive tags that characterize the transactions in the retail bank dataset. For evaluation, 12 volunteers completed a two-part form assessing the quality of the taxonomies and the tags assigned to merchants. The evaluation revealed a coherence rate exceeding 90% for the chosen taxonomies. Additionally, taxonomy expansion using LLMs demonstrated promising results for parent node prediction, with F1-scores of 89% and 70% for Food and Shopping taxonomies, respectively.

This work presents an unsupervised method for tagging banking consumers’ transactions using automatically constructed and expanded topic taxonomies. Initially, we enrich the bank transactions via web scraping to collect relevant descriptions, which are then preprocessed using NLP techniques to generate candidate terms. Topic taxonomies are created using instruction-based fine-tuned LLMs (Large Language Models). To expand existing taxonomies with new terms, we use zero-shot prompting to determine where to add new nodes. The resulting taxonomies are used to assign descriptive tags that characterize the transactions in the retail bank dataset. For evaluation, 12 volunteers completed a two-part form assessing the quality of the taxonomies and the tags assigned to merchants. The evaluation revealed a coherence rate exceeding 90% for the chosen taxonomies. Additionally, taxonomy expansion using LLMs demonstrated promising results for parent node prediction, with F1-scores of 89% and 70% for Food and Shopping taxonomies, respectively.

Substantial efforts have been dedicated to developing methods for detecting deepfake content, especially with the creation of large and diverse datasets with both higher image quality and demographic features. In this scenario, CNN-based approaches showed good initial success, later improved by their combination with Vision Transformers. More recently, Foundation Models (FMs) have emerged, improving performance across many visual tasks, including deepfake detection, and combining self-supervised features generated by FMs with CNN-based classifiers has resulted in significant performance gains. However, taking advantage of multiple maps of self-supervised features is not as straightforward as just adding more channels to the classifier. Therefore, this work explores ensemble techniques to effectively utilize these diverse self-supervised feature maps for realistic facial deepfake detection. Our experiments indicate that combining the output results of different classifiers, each one utilizing a single map of self-supervised features, leads to significant performance improvements, and several committee approaches consistently outperform individual classifiers, demonstrating the potential of these methods in enhancing deepfake detection accuracy.

This paper explores potential applications for TV 3.0, a ground-breaking evolution in Brazilian digital television that introduces advanced functionalities such as app-based experience, viewer-centered design, enhanced interactivity, and integration with smart technologies. Leveraging these new capabilities, TV 3.0 can revolutionize content delivery and audience engagement in several key areas by enabling personalized content experiences with recommendations and adaptive streaming to enhance viewer satisfaction and retention. We present three use cases that explore some of the new features and their potential impacts on content transmission and consumption and the development of interactive applications.