Publications

You can also find my articles on my Google Scholar profile.

Books

Normative Ethics, Artificial Intelligence, and Value Alignment: From Normative Theory to Engineering Practices

Published in Springer Nature, 2026

The book examines value alignment from the bottom up. First, we lay the philosophical roots of the project, encapsulated in what I call Dynamic Normativity, and later we move to implementation, showing how this theory would be put into practice. The whole journey takes the reader from metaphysics and metaethics to the intricacies of post-training language models.

Standardization of AI Ethics: Stakeholders, Values and Profit

Published in transcript Verlag, 2026

This book explores the complex relationship between AI, ethics, societal values and standardization, discussing the challenges and opportunities involved in standardizing AI ethics, including algorithmic audits, stakeholder participation and certification mechanisms.

Dynamic Normativity: Necessary and Sufficient Conditions for Value Alignment

Published in Doctoral Dissertation, 2024

This work reframes AI alignment as a technical‑philosophical problem requiring both solid normative foundations and practical implementation. It introduces Dynamic Normativity, a framework built on necessary and sufficient conditions for value‑aligned learning systems, and demonstrates how these principles can guide the alignment of modern language models.

Journal Articles

From computation to environmental cost: the resource burden of artificial intelligence

Published in Nature: Communications Earth & Environment, 2026

In this study, we quantify the material footprint of artificial intelligence training by linking computational workloads to physical hardware needs. We analyze the elemental composition of a widely used graphics processing unit and find that AI hardware consists of approximately 90% heavy metals. Training a large language model requires between 1760 and 8800 graphics processing units, highlighting the need to incorporate material resource considerations into discussions of AI scalability and sustainability.

Reasoning Primitives in Hybrid and Non-Hybrid LLMs

Published in ArXiV, 2026

This paper explores reasoning primitives in hybrid and non-hybrid large language models, focusing on recall and state-tracking capabilities. We evaluate matched Olmo3 transformer and hybrid models on tasks that require both primitives, finding that reasoning augmentation extends the effective operating range of models. The hybrid model shows greater robustness to increasing sequential dependence, while the transformer model degrades sharply beyond a certain difficulty threshold.

LilTii: A 0.6B Bengali Language Model that Outperforms Qwen

Published in Hugging Face Blog, 2026

This blog post introduces LilTii, a 0.6B-parameter Bengali language model trained from scratch to address the gap in NLP resources for low-resource languages. We compiled a high-quality Bengali corpus using heuristic and learned filtering, supplemented with curated English data for bilingual augmentation. LilTii consistently outperforms similarly sized multilingual models such as Qwen2.5-0.5B and Qwen3-0.6B across various Bengali benchmarks, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.

Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi

Published in ArXiV, 2026

This study introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address the gap in NLP resources for low-resource languages. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned methods, complemented by bilingual augmentation with curated English data. LilMoo consistently outperforms comparably sized multilingual baselines, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.

Tucano 2 Cool: Better Open Source LLMs for Portuguese

Published in ArXiV, 2026

This study introduces Tucano 2, a suite of open-source large language models (LLMs) with 0.5-3.7 billion parameters, designed to fill gaps in Portuguese LLM development. We extend our dataset, GigaVerbo-v2, and introduce new synthetic and post-training datasets to enhance training across various domains. Through extensive ablation studies, we develop pretraining and continual pretraining recipes that achieve state-of-the-art performance on Portuguese-language benchmarks.

From FLOPs to Footprints: The Resource Cost of Artificial Intelligence

Published in ArXiV, 2025

This study quantifies the material footprint of AI training by linking computational workloads to the elemental composition of modern GPUs. By showing how model scale drives heavy‑metal extraction and hardware turnover, it demonstrates that AI progress carries substantial material costs and argues for integrating resource efficiency into future discussions of AI scalability.

Tucano: Advancing neural text generation for Portuguese

Published in Patterns, 2025

This study introduces GigaVerbo, a 200‑billion‑token Portuguese corpus, and Tucano, a family of transformer models trained on it. By addressing the resource gap in neural text generation, these models outperform comparable Portuguese and multilingual systems and are released openly to foster reproducible research.

Technical, legal, and ethical challenges of generative artificial intelligence: an analysis of the governance of training data and copyrights

Published in Discover Artificial Intelligence, 2025

This study analyzes the legal and ethical tensions surrounding generative AI, emphasizing how opaque training data practices undermine copyright compliance and accountability. By comparing regulatory approaches across major jurisdictions, it shows why existing mechanisms remain insufficient and proposes transparency mandates, compensation schemes, and audit tools as a path toward fair and sustainable AI governance.

Codes of ethics in IT: do they work in isolation?

Published in AI and Ethics, 2025

This study evaluates whether exposure to professional codes of ethics meaningfully shapes ethical decision‑making in IT. Using a randomized controlled trial with a multimedia version of the ACM Code of Ethics, it finds no measurable behavioral impact from passive exposure and argues for exploring more active, engaged forms of ethical intervention in future research.

Superalignment With Dynamic Human Values

Published in ICLR 2025 Workshop on Bidirectional Human-AI Alignment (BiAlign), 2025

This work proposes a new alignment framework that unifies scalable oversight with the evolving nature of human values. By training a superhuman reasoning model to decompose tasks and leveraging the part‑to‑complete generalization hypothesis, it outlines a path for ensuring that aligned subtask solutions reliably extend to full, complex behaviors.

Catalog of General Ethical Requirements for AI Certification

Published in Whitepaper, 2024

This whitepaper outlines a practical and normative framework for achieving Trustworthy AI, pairing six core ethical principles with concrete, value‑specific implementation tools. By linking fairness, privacy, safety, sustainability, transparency, explainability, and truthfulness to risk‑assessment criteria aligned with the EU AI Act, it offers stakeholders a blueprint for meeting minimum ethical and certification requirements.

TeenyTinyLlama: Open-source tiny language models trained in Brazilian Portuguese

Published in Machine Learning with Applications, 2024

This study presents TeenyTinyLlama, a pair of compact, open‑foundation language models designed for Brazilian Portuguese. By addressing the limitations of multilingual LLMs in low‑resource settings, it demonstrates how smaller, permissively licensed models can offer practical, community‑driven alternatives for localized text generation.

Crossing the principle–practice gap in AI ethics with ethical problem-solving

Published in AI and Ethics, 2024

This work addresses the growing gap between AI ethics principles and their practical implementation. It introduces ethical problem‑solving (EPS), a framework that translates high‑level values into actionable design choices through impact assessments and differential recommendations, offering a concrete path toward responsible, human‑centric AI development.

Worldwide AI ethics: A review of 200 guidelines and recommendations for AI governance

Published in Patterns, 2023

This paper analyzes 200 global AI governance policies to identify common ethical principles guiding responsible AI use. By mapping recurring concerns—privacy, discrimination, security, transparency, and other systemic risks—it reveals at least 17 shared principles and evaluates their significance for future regulatory frameworks.

Counterfactual Analysis by Algorithmic Complexity: A metric between possible worlds

Published in Manuscrito, 2022

This study introduces algorithmic complexity as a new method for analyzing counterfactuals. Building on Lewis-Stalnaker’s possible worlds semantics, it offers fresh insights into debates on vagueness, context-dependence, and non-monotonicity, making counterfactual reasoning more intuitive and philosophically robust.

On the Efficiency of Ethics as a Governing Tool for Artificial Intelligence: A Critical Comparison between AI Ethics and Bioethics

Published in ArXiV, 2022

This paper critically examines AI Ethics as a governance tool, arguing that principled guidelines alone are insufficient. It calls for stronger regulation, revised professional training, and integration of bioethics into law to ensure safer, more accountable AI development.

Technical Note on the Risks of Using Facial Recognition Technologies in Public Security Applications

Published in Whitepaper, 2022

This note examines the expansion of facial recognition in Porto Alegre and outlines the major risks of deploying FRTs for public safety. By integrating technical, legal, and ethical analysis, it clarifies how these systems work, where they fail, and why their use demands careful public scrutiny.

Good AI for the Present of Humanity: Democratizing AI Governance

Published in The AI Ethics Journal, 2021

This article connects Cyberpunk and AI Ethics, arguing for a more democratic approach to AI governance. By exposing deficits in current industry power structures, it advocates that public opinion should guide ethical frameworks so “good AI” becomes “good AI for all.”

Singularity and Coordination Problems: Pandemic Lessons from 2020

Published in Journal of Futures Studies, 2021

This paper reviews safety research in AI and examines paradigms that may signal a technological singularity. Using the COVID-19 pandemic as a case study, it highlights humanity’s lack of global coordination in managing existential risks.

Metanormativity: Solving questions of moral and empirical uncertainty

Published in ethic@-An international Journal for Moral Philosophy, 2020

This study explores metanormativity (norms about norms) as a framework for resolving moral uncertainty. Using vegetarian dilemmas, it applies William MacAskill’s Maximization of Expected Choice-Worthiness, showing how ethical-mathematical models can aggregate conflicting moral theories and guide decision-making under uncertainty.

Dynamic Models Applied to Value Learning in Artificial Intelligence

Published in ArXiV, 2020

This paper discusses the challenge of aligning artificial intelligence with human values. It highlights risks of uncritical development, critiques symbolic and connectionist approaches, and proposes dynamic, phenomenological models, such as situated embodied dynamics (SED), as a pathway for safer, value-sensitive AI.

Development of a skateboarding trick classifier using accelerometry and machine learning

Published in Research on Biomedical Engineering, 2017

This study developed machine learning classifiers for skateboarding tricks using inertial measurement unit data. Artificial neural networks trained on modeled acceleration signals achieved error rates below 0.05%, demonstrating the potential for real-time, accurate trick detection in street skating.

The evaluation of physiological responses during the performance of external chest compressions in simulated hypo and microgravity

Published in Resuscitation, 2012

This study examined CPR performance under Earth gravity, hypogravity, and microgravity. While compression quality remained consistent, rescuers showed higher heart rates, exertion, and altered muscle activation in reduced gravity, indicating greater physical effort is required for effective CPR during space missions.

Muscle Activity during the Performance of CPR in Simulated Microgravity and Hypogravity

Published in American Journal of Medical and Biological Research, 2012

This study compared CPR performance in Earth gravity, hypogravity, and microgravity. While compression rates were consistent, reduced gravity increased heart rate, exertion, and muscle activation, especially in the rectus abdominis and pectoralis major. Findings highlight the need for strong physical conditioning to ensure effective CPR in extraterrestrial environments.

Gender Influence on the Performance of Chest Compressions in Simulated Hypogravity and Microgravity

Published in Aviation, Space and, Environmental Medicine, 2012

In the event of a cardiac arrest during microgravity exposure, external chest compressions (ECCs) which form the main part of basic life support should be carried out while the advanced life support equipment is being deployed. This study was aimed to determine if there was any gender difference in the effectiveness of performing ECCs using a body suspension device to simulate lunar and Martian hypogravity and microgravity.