GeezerVison is showing uffescience.ic.uff.br — UFFeScience Research Group — Projects and research
UNIVERSIDADE FEDERAL FLUMINENSE
UFFeScience Research Group
Welcome to the official page of the UFF eScience Research Group, part of the Institute of Computing at Universidade Federal Fluminense (UFF).
Our group is dedicated to advancing research in Data Science, Data Management, Machine Learning, and their applications in multiple domains of Science (Bioinformatics, Astronomy, etc). Here you will find information about our team, ongoing projects, and scientific publications.
🎤 UFFeScience Research Group at SSCAD 2026!
Prof. Daniel de Oliveira was in Natal this week for the Simpósio em Sistemas Computacionais de Alto Desempenho (SSCAD 2026), where he presented our paper:
📄 “Can Large Language Models Engineer Parallel Workflows?”
Authors: Paula W., Liliane Kunstmann, Débora Pina, Marta Mattoso, Daniel de Oliveira
LLMs are increasingly used to write code, but can they build correct parallel scientific workflows? In this paper, we evaluate how well LLMs compose workflows for parallel workflow systems such as Parsl, PyCOMPSs, and Nextflow, and discuss where they succeed and where human expertise still matters.
The paper is part of Paula Woyames’s Master’s research, and we’re already working on the next steps of this line of work. Stay tuned! 🚀
Thanks to the SSCAD organizers and to everyone who joined the session for the great discussion.
🔗 Read the paper: https://lnkd.in/dekjEUjj
#SSCAD2026 #HPC #LLM #ScientificWorkflows #ParallelComputing #AI #eScience #UFF
Can LLMs recognize ableism? Our latest work at KDMiLe! ⛏️📊
Another paper from the UFFeScience Research Group, in collaboration with MeLLL@UFF, at KDMiLe - Knowledge Discovery, Mining and Learning 2026!
Yesterday in Cuiabá, Gabriel Assis presented our paper “Beyond Toxicity: Evaluating Language Models for Ableism Detection in Online Social Media” at KDMiLe 2026.
Ableism on social media isn’t always explicit. It often shows up as stereotypes, jokes, and comments that traditional toxicity detection approaches miss.
This paper is part of Daniele Pimenta's undergrad final project and analyzes how LLMs handle this task when given more context about the events being analyzed, one piece at a time.
👥 Authors: Annie Amorim, Gabriel Assis, Daniele Pimenta, Débora Pina, Aline Paes, and Daniel de Oliveira.
Congratulations to the whole team! 👏
📄 Read the paper: https://lnkd.in/de9Fh63e
#KDMiLe #LLMs #NLP #Ableism #Accessibility #ArtificialIntelligence #UFF #UFFeScience #MeLLL
🌊 AkôFlow is heading to Tomorrow.Blue Economy 2026!
We are happy to share that our student, Wesley Ferreira, had his talk proposal selected in the AGIR/PROPPI–UFF internal call for proposals for Tomorrow.Blue Economy, taking place November 25–26, 2026, at Arena Niterói.
Wesley will present AkôFlow, the open-source engine developed by the UFFeScience Research Group for running containerized workflows across the computing continuum, from edge devices to the cloud.
Data-intensive problems, including those at the heart of the ocean economy, increasingly depend on infrastructure that can move computation to where the data lives. We are excited to bring this discussion to one of the world's largest blue economy events, right here in Niterói.
Congratulations, Wesley! 👏
And many thanks to AGIR and PROPPI/UFF for the opportunity.
🔗 AkôFlow on GitHub: https://lnkd.in/eqdx_3iK
#AkôFlow #BlueEconomy #TomorrowBlueEconomy #eScience #UFF #Workflows #CloudComputing #EdgeComputing #OpenSource #Niterói
📢 Paper accepted at SSCAD 2026!
We are happy to announce that our paper “Can Large Language Models Engineer Parallel Workflows?” has been accepted at SSCAD 2026! 🎉
The paper is by Paula W. (IC/UFF), Liliane Kunstmann (UFRRJ), Débora Pina (Barcelona Supercomputing Center), Marta Mattoso (COPPE/UFRJ), and Daniel de Oliveira (IC/UFF).
Developing a workflow that actually runs in parallel requires knowing the specification language and execution model of a specific workflow system, and many domain scientists don’t have that expertise. LLMs promise to lower this barrier, but existing evaluations mostly stop at whether the generated code looks right. We wanted to know whether it runs, and whether it runs in parallel.
We evaluated four LLMs (GPT-5.5, DeepSeek-V4-Flash, Command-A-Plus, and Gemma-4-31B-IT) on two tasks:
Specification: writing a workflow from a natural-language description.
Conversion: translating an existing workflow from one system to another.
The study covered three workflow systems (Parsl, PyCOMPSs, and Nextflow), three workflows (K-Means, MatMul, and a GROMACS molecular dynamics simulation), and zero-shot and few-shot prompting. We executed all 600 generated specifications in a real environment.
The takeaway is that LLMs can help with workflow engineering, but evaluations that only check whether code runs overestimate how useful the output really is in HPC and cloud environments.
Congratulations to Paula and to all co-authors! 👏
💻 Prompts, specifications, and scripts: https://lnkd.in/dpCBrkUV
#UFFeScience #UFF #SSCAD2026 #LLM #ScientificWorkflows #HPC #ParallelComputing #Parsl #PyCOMPSs #Nextflow #AI4Science
We are proud to announce that Marcello Ribeiro successfully defended his Ph.D. thesis, “An Approach for Scheduling Scientific Workflows in Hybrid VM–Serverless Cloud Environments,” on October 6 and is now the newest Doctor at the Instituto de Computação at Universidade Federal Fluminense (IC/UFF).
Workflows increasingly execute on elastic cloud infrastructures, where serverless functions now coexist with traditional Virtual Machines. Each model has its own performance, pricing, and operational constraints, so a scheduler must decide not only when each activation runs, but also where. Marcello’s thesis tackles this underexplored heterogeneous scheduling problem with three complementary contributions:
🔹 Denethor, a provenance capture system for workflows running on serverless platforms, providing empirical execution time and cost data;
🔹 FLEXE+IP, an integer programming model that formulates the scheduling problem;
🔹 FLEXE+, a GRASP-based heuristic with VND local search that jointly chooses the execution environment and the execution order, balancing makespan and financial cost.
Experiments with synthetic workflows and a large-scale real bioinformatics workflow show that FLEXE+ reduces makespan or cost depending on the prioritized objective, while reaching near-optimal solutions on instances where the exact model is tractable.
Marcello, thank you for your dedication and for all the contributions you brought to our group over these years. Well deserved!
Marcello was advised by Profs. Daniel de Oliveira and (UFF) Isabel Rosseti (UFF). Our thanks to the examiners for the insightful discussion: Simone Martins, Yuri Frota (UFF), Marta Mattoso (COPPE/UFRJ), Luan Teylo (INRIA), Rafaelli Coutinho (Cefet/RJ), and UBIRATAM CARVALHO DE PAULA JUNIOR (UFRRJ).
#PhD #CloudComputing #Serverless #ScientificWorkflows #Scheduling #Metaheuristics #UFF #eScience
📢 New paper at SIBGRAPI 2026!
We're happy to share "Reveal: A Provenance-Aware System for Multimodal Urban Data Exploration," presented at SIBGRAPI 2026.
Reveal is a visual analytics system that tackles:
🔎 Multimodal text and image search over a shared CLIP embedding space
🧩 Set operations to combine and refine urban concepts
🗺️ Linked views: embedding plots, image galleries, word clouds, and spatial heatmaps
🧭 An exploration diagram that captures analytical provenance, making each analysis transparent, reproducible, and reusable
In case studies with an urban planner, Reveal helped expand the Project Sidewalk accessibility taxonomy with 11 new fine-grained categories, covering more than 1,600 newly labeled images.
We then reused the captured provenance as a "semantic blueprint" to replicate the analysis on Chicago street-level imagery.
This work continues our research line on provenance and scientific data management, applied here to urban visual analytics. It is a collaboration between the UFFeScience Research Group (Daniel de Oliveira, Aline Paes, and Marcos Lage), UIC (Leonardo Ferreira and Fabio Miranda), FGV (Joao Rulff), and UC Berkeley (Maryam Hosseini), and it builds on Leonardo Ferreira's master's work.
Congratulations to all co-authors!
💻 Code and tutorial: https://lnkd.in/dGzXh5V9
#SIBGRAPI2026 #eScience #UFF #Provenance #VisualAnalytics #UrbanComputing #MultimodalData
On September 17, Prof. Aline Paes (MeLLL@UFF) and Prof. Daniel de Oliveira (UFFeScience Research Group) had the pleasure of giving a talk at CEFET/RJ – Campus Maria da Graça.
They presented "Finding the Unexpected: Machine Learning for the Next Generation of Astronomy," based on Camila Lopes's PhD research.
Modern telescopes produce so much data that analyzing everything with traditional methods is becoming impossible. In the talk, they showed how machine learning can help astronomers find unusual and potentially important objects hidden in huge datasets. Their focus was Federated Learning, an approach that lets different computers train models together without moving all their data to one central location. Using observations from the Dark Energy Survey (DES), they tested how this approach could support future projects like the Legacy Survey of Space and Time (LSST), comparing configurations by speed, reliability, and cost.
Huge thanks to the CEFET/RJ students for their great questions and genuine curiosity. It's inspiring to see young people so interested in our work!
Thanks also to Prof. Carlos Eduardo Pantoja and CEFET/RJ for hosting us and welcoming us so warmly.
#MachineLearning #FederatedLearning #Astronomy #DataScience #eScience #UFF #CEFETRJ
Happy to share that our paper entitled "A Workflow for Automated Cell-Type Annotation in scRNA-seq Using LLMs and Graph-Based Community Detection" has been accepted at the Brazilian Symposium on Bioinformatics (BSB 2026)!
The paper proposes CellLens+, which is a workflow for automated cell-type annotation that brings together graph-based community detection, deterministic molecular inference, and Large Language Models. It detects transcriptionally coherent cell populations, identifies cluster-specific marker genes, and assigns cellular identities through a quantitative scoring framework.
Authors: Layse Gomes Castro, João Vitor Moraes, and Daniel de Oliveira.
Congratulations, and see you at BSB 2026 in Ribeirão Preto!
#Bioinformatics #SingleCell #scRNAseq #LLM #BSB2026 #UFF
Congratulations to Valesca Moura, who presented the work "Instrumentation vs. Observability: Provenance Capture Trade-offs in Agentic Workflows" at the APeX Summit 2026 (Academy, Parsl, Globus Compute).
As AI agents increasingly take part in scientific workflows, understanding what they did, why, and with which data becomes important for trust and reproducibility. This work looks at a practical question at the heart of this: when capturing provenance in agentic workflows, what do we gain and what do we pay when we instrument the code explicitly versus relying on observability mechanisms?
And there's more good news! The paper has been accepted at WISDOM @ IEEE eScience 2026 and will be published soon.
This was a truly collaborative effort, and we are grateful to work with such a great team: Valesca Moura, Amal Gueroudji, Daniel Rosendo, Ben Clifford, Kyle Chard, Rafael Ferreira da Silva, Daniel de Oliveira and Renan Souza.
Looking forward to what comes next.
🔗 Program and slides: https://lnkd.in/d9c5draW
#eScience #Provenance #AgenticAI #ScientificWorkflows #Observability
🇧🇷 pt-br
Mais uma boa notícia direto do Brazilian Symposium on Databases (SBBD): nosso artigo recebeu o prêmio de Melhor Artigo do XX Brazilian e-Science Workshop (BreSci 2026)! 🏆
"Mineração de Subárvores Frequentes em Árvores Filogenéticas usando Conjuntos de Itens Frequentes Maximais" é fruto do trabalho de João Vitor Moraes e Layse Gomes Castro, em parceria com Isabel Rosseti e Daniel de Oliveira, todos do Instituto de Computação UFF.
O ponto de partida é que métodos diferentes produzem árvores filogenéticas distintas a partir dos mesmos dados, e as árvores de consenso — a saída usual — acabam por esconder conflitos topológicos.
O PhyloTreeMiner propõe uma alternativa ao consenso: minerar subárvores frequentes, identificando padrões topológicos recorrentes com base no valor de suporte. Modelamos o problema como mineração de conjuntos de itens frequentes maximais (FPMax), o que reduz o custo computacional, e armazenamos os resultados no Neo4j, permitindo consultas que combinam topologia e metadados.
Parabéns, João Vitor e Layse — o prêmio é de vocês! 👏
📄 Artigo: https://lnkd.in/ddQNTUFB
#BreSci2026 #eScience #Filogenia #Bioinformática #DataMining #Neo4j #UFF
🇺🇸 en-us
More good news from Brazilian Symposium on Databases (SBBD): our paper received the Best Paper Award at the XX Brazilian e-Science Workshop (BreSci 2026)! 🏆
"Frequent Subtree Mining in Phylogenetic Trees using Maximal Frequent Itemsets" is the work of João Vitor Moraes and Layse Gomes Castro, together with Isabel Rosseti and Daniel de Oliveira, all from IC/UFF.
Different methods produce different phylogenetic trees from the same dataset, and consensus trees—the usual way out—tend to hide topological conflicts and discard relevant evolutionary signal.
PhyloTreeMiner offers an alternative to consensus: mining frequent subtrees to identify recurring topological patterns based on support values. We model the problem as maximal frequent itemset mining (FPMax), which substantially reduces the computational cost, and store the results in Neo4j, enabling queries that combine topology and metadata.
Congratulations, João Vitor and Layse — this award is yours! 👏
📄 Paper: https://lnkd.in/ddQNTUFB
#BreSci2026 #eScience #Phylogenetics #Bioinformatics #DataMining #Neo4j #UFF
🇧🇷 pt-br
O UFFeScience Research Group desembarcou em São Carlos para o Brazilian Symposium on Databases (SBBD) 2026.
Foram apresentados 11 trabalhos, entre artigos completos, artigos curtos, demonstrações e workshops, abrangendo temas como proveniência de workflows agênticos, IA responsável, arquiteturas HTAP, continuum computacional, grafos de conhecimento e bioinformática.
E dois reconhecimentos que coroaram a semana:
🏆 Melhor Artigo do BreSci 2026
🏅 Menção Honrosa de artigos completos no SBBD 2026
Mas o número que mais importa não é nenhum desses, é o nosso grupo. Alunos e alunas que transformaram noites de experimento em resultado, coautores e parceiros de outras instituições que somaram conosco, e uma organização impecável do SBBD e do BreSci que fez tudo acontecer.
Obrigado a cada um de vocês. 👏
Os artigos estão todos nos comentários. 👇
#SBBD2026 #BreSci2026 #eScience #UFF #BancoDeDados
🇺🇸 en-us
The UFFeScience Research Group landed in São Carlos for Brazilian Symposium on Databases (SBBD) 2026.
Eleven works were presented as full papers, short papers, a demo, and workshop contributions, covering topics such as provenance for agentic workflows, responsible AI, HTAP architectures, the computing continuum, knowledge graphs, and bioinformatics.
Plus two recognitions that capped the week:
🏆 Best Paper Award at BreSci 2026
🏅Full Paper Honorable Mention at SBBD 2026
But the number that matters most isn't any of those, it's our team. The students who turned late-night experiments into results, the co-authors and partners from other institutions who joined us, and the SBBD and BreSci organizers who made it all run flawlessly.
Thank you, every one of you 👏
All the papers are in the comments. 👇
#SBBD2026 #BreSci2026 #eScience #UFF #Databases
🇧🇷 pt-br
Notícia que dá orgulho de compartilhar: o artigo associado a pesquisa de mestrado da Valesca Moura (orientada pelos Profs. Sandro Rama Fiorini e Daniel de Oliveira) recebeu Menção Honrosa no SBBD 2026!
O artigo "Uma Abordagem Orientada a Dados de Proveniência para Rastreabilidade de Workflows Agênticos" aborda um problema que cresce com a adoção de agentes de IA: quando um LLM decide as tarefas e as dependências de dados de um workflow, como saber por que o workflow ficou daquele jeito?
A abordagem (ANCHOR) captura a proveniência de workflows agênticos, instrumentando o próprio framework de agentes e representando os dados conforme o padrão W3C PROV. O resultado é a possibilidade de analisar as decisões dos agentes e o impacto delas no workflow gerado, com granularidade superior à dos trabalhos existentes.
Parabéns, Valesca! 👏
📄 Artigo: https://lnkd.in/dT2ra9aC
💻 Código: https://lnkd.in/dsSVpD-U
#SBBD2026 #Proveniência #WorkflowsAgênticos #AgenticAI #IAResponsável #UFF #eScience
🇺🇸 en-us
Proud to share this one: Valesca Moura's master's research (advised by Sandro Fiorini and Daniel de Oliveira) received an Honorable Mention at SBBD 2026! 🏆
The paper "A Provenance Data-Oriented Approach for Agentic Workflow Traceability" tackles a problem that grows with every new AI agent deployment: when an LLM decides tasks and data dependencies, how do we know why a workflow ended up the way it did?
Our approach (ANCHOR) captures provenance from agentic workflows by instrumenting the agent framework itself and representing the data according to the W3C PROV standard. This makes it possible to analyze agents' decisions and their impact on the generated workflow, at a finer granularity than existing approaches offer.
Congratulations, Valesca! 👏
📄 Paper: https://lnkd.in/dT2ra9aC
💻 Code: https://lnkd.in/dsSVpD-U
#SBBD2026 #DataProvenance #AgenticWorkflows #AgenticAI #ResponsibleAI #UFF #eScience
🇧🇷 Nova publicação do UFFeScience Research Group!
Temos o prazer de anunciar a publicação do artigo "Optimizing Preventive Police Patrol Routes in Brazilian Cities Using a GRASP-Based Routing Heuristic" no Journal of Internet Services and Applications (JISA).
O trabalho aborda o problema de Roteamento de Viaturas Policiais em Grandes Centros Urbanos (RVP-Urb), que busca maximizar a cobertura de áreas críticas de criminalidade respeitando restrições de recursos. A solução proposta, chamada PROTECTOR, é uma heurística baseada no GRASP (Greedy Randomized Adaptive Search Procedure) para gerar rotas de patrulhamento preventivo que aumentam a visibilidade policial em hotspots criminosos.
👥 Autores: Isabel Rosseti, Eduardo Canellas, Wagner Santos, Marcos Lage, Yuri Frota e Daniel de Oliveira.
Parabéns a toda a equipe!
🔗 DOI: https://lnkd.in/d2X_giC3
#UFFeScience #UFF #InstitutoDeComputação #JISA #Otimização #GRASP #Metaheurísticas #CidadesInteligentes #CiênciaDeDados #Pesquisa
🇺🇸 New publication from UFFeScience Research Group!
We're glad to announce the publication of "Optimizing Preventive Police Patrol Routes in Brazilian Cities Using a GRASP-Based Routing Heuristic" in the Journal of Internet Services and Applications (JISA).
The paper tackles the Routing of Police Vehicles in Large Urban Centers (RVP-Urb) problem, which aims at maximizing crime coverage under resource constraints. The proposed solution, PROTECTOR, is a heuristic based on the Greedy Randomized Adaptive Search Procedure (GRASP) that generates preventive patrol routes to enhance police visibility across crime hotspots.
👥 Authors: Isabel Rosseti, Eduardo Canellas, Wagner Santos, Marcos Lage, Yuri Frota, and Daniel de Oliveira.
Congratulations to the whole team!
🔗 DOI: https://lnkd.in/d2X_giC3
#UFFeScience #UFF #JISA #Optimization #GRASP #Metaheuristics #SmartCities #DataScience
🇧🇷 Mais um trabalho em cooperação com o MeLLL@UFF!
Nosso artigo "Beyond Toxicity: Evaluating Language Models for Ableism Detection in Online Social Media" foi aceito para apresentação no KDMiLe - Knowledge Discovery, Mining and Learning 2026, em Cuiabá.
O artigo investiga como LLMs identificam manifestações de capacitismo em comentários em redes sociais em português brasileiro. Avaliamos diferentes estratégias de prompting e mostramos que contexto adicional pode melhorar o desempenho dos modelos, embora ainda existam desafios importantes na distinção entre capacitismo, críticas ao capacitismo e outros tipos de discurso discriminatório.
Parabéns a Annie Amorim, Gabriel Assis, Daniele Pimenta, Débora Pina, Aline Paes e Daniel de Oliveira!
Nos vemos no KDMiLe 2026. 🎉
#LLM #AI #NLP #ResponsibleAI #Research #KDMiLe2026
🇺🇸 Another publication in cooperation with MeLLL@UFF!
Our paper, "Beyond Toxicity: Evaluating Language Models for Ableism Detection in Online Social Media", has been accepted for presentation at KDMiLe - Knowledge Discovery, Mining and Learning 2026 in Cuiabá, Brazil.
The study investigates how LLMs identify ableism in Brazilian Portuguese social media comments. We evaluated different prompting strategies and found that additional context can improve model performance, although important challenges remain in distinguishing ableism from criticism of ableism and other forms of discriminatory discourse.
Congratulations to Annie Amorim, Gabriel Assis, Daniele Pimenta, Débora Pina, Aline Paes, and Daniel de Oliveira!
See you at KDMiLe 2026! 🎉
#LLM #ArtificialIntelligence #NLP #ResponsibleAI #Research #KDMiLe2026
🎉 UFFeScience Research Group no SBBD 2026!
O UFFeScience Research Group (Instituto de Computação UFF) teve 8 artigos aceitos no XLI SBBD Simpósio Brasileiro de Banco de Dados: 4 completos e 4 curtos! E dois deles concorrem aos prêmios de Melhor Artigo Completo e Melhor Artigo Curto 🏆
📄 ARTIGOS COMPLETOS
Uma Abordagem Orientada a Dados de Proveniência para Rastreabilidade de Workflows Agênticos (concorre ao melhor full paper 🏆)
Valesca Moura, Sandro Rama Fiorini, Daniel de Oliveira
🔗 https://lnkd.in/dtjmPKMi
Triptych: Integrando Dados, Código e Publicações para Pesquisa Científica
Ygor Rolim, Liliane Kunstmann, Débora Pina, Carina F. Dorneles, Marcos Lage, Daniel de Oliveira
🔗 https://lnkd.in/dJ4gUtKM
Do Data Warehouse à Narrativa: Uma Arquitetura HTAP para Análise de Dados de Criminalidade
Ryan Serra, Daniel Ribeiro, Arben Hajrullah S. Tafili, Paulo S. Costa-Jr., João V. Silva-Leite, Rodrigo Salvador Monteiro, Daniel de Oliveira, Marcos Bedo
🔗 https://lnkd.in/d9SQtPu2
Definição de Acordos de Nível de Serviço Personalizados para Treinamento de Modelos via Consultas k-Skyband
Gustavo Alexandre Santos, Marcos Bedo, Yuri Frota, Daniel de Oliveira
🔗 https://lnkd.in/dYF39E-t
📄 ARTIGOS CURTOS
Understanding Storage Trade-offs for Containerized Dataflows Across the Computing Continuum (concorre ao melhor short paper 🏆)
Wesley Ferreira, Luiz Viana, Lucas Natan, Liliane Kunstmann, Daniel de Oliveira
🔗 https://lnkd.in/dFHXaWhA
On Provenance in Model Fusion
Annie Amorim, João Vitor Moraes, Débora Pina, Aline Paes, Daniel de Oliveira
🔗 https://lnkd.in/d6FDfx-i
Provenance Graph Integration in Scientific Machine Learning Multi-workflows
Débora Pina, Lucas Moreira Dias, Maria Célia Santos Lopes, Daniel de Oliveira, Marta Mattoso
🔗 https://lnkd.in/dZBvNywZ
Rumo a Duas Décadas de Publicações em Banco de Dados: Uma Análise dos Artigos do JIDM
Elena Costa Blanco, Josué Santos, Maristela Holanda, Vanessa Braganholo, Marcos Bedo, Daniel de Oliveira
🔗 https://lnkd.in/dJdxheEz
Um obrigado especial aos nossos alunos, da iniciação científica ao doutorado: a dedicação de vocês, nas madrugadas de experimento e nas revisões, está em cada página desses artigos. Vocês são o motor deste grupo 💙
Parabéns a todas e todos! Nos vemos em São Carlos!
#SBBD2026 #UFF #eScience #BancoDeDados
🇧🇷 (pt-br) 🎉 Ótimas notícias para começar o fim de semana!
É com muita alegria que compartilhamos que dois trabalhos com participação do UFFeScience Research Group foram indicados a prêmios no Brazilian Symposium on Databases (SBBD) 2026, principal evento da comunidade de bancos de dados no Brasil.
🏅 Indicação a Melhor Artigo Curto
"Understanding Storage Trade-offs for Containerized Dataflows Across the Computing Continuum". Autores: Wesley Ferreira (IC/UFF), Luiz Viana (IC/UFF), Lucas Natan (IC/UFF), Liliane Kunstmann (UFRRJ) e Daniel de Oliveira (IC/UFF).
🏅 Indicação a Melhor Artigo Completo (Best Full Paper)
"Uma Abordagem Orientada a Dados de Proveniência para Rastreabilidade de Workflows Agênticos". Autores: Valesca Moura (IC/UFF), Sandro Rama Fiorini (Instituto PUC-Behring de Inteligência Artificial) e Daniel de Oliveira (IC/UFF).
Parabéns a todos os autores e autoras! Indicações como essas são resultado de meses de dedicação, muitas iterações e de um trabalho verdadeiramente coletivo.
Nos vemos no SBBD!
🇺🇸 (en-us) 🎉 Great news to kick off the weekend!
We're delighted to share that two papers co-authored by the UFFeScience Research Group have been nominated for awards at the Brazilian Symposium on Databases (SBBD) 2026, the leading venue of the Brazilian database community.
🏅 Best Short Paper Nomination
"Understanding Storage Trade-offs for Containerized Dataflows Across the Computing Continuum". Authors: Wesley Ferreira (IC/UFF), Luiz Viana (IC/UFF), Lucas Natan (IC/UFF), Liliane Kunstmann (UFRRJ), and Daniel de Oliveira (IC/UFF).
🏅 Best Full Paper Nomination
"A Provenance-Driven Approach for the Traceability of Agentic Workflows". Authors: Valesca Moura (IC/UFF), Sandro Rama Fiorini (Instituto PUC-Behring de Inteligência Artificial), and Daniel de Oliveira (IC/UFF).
Congratulations to all the authors! Nominations like these are the result of months of dedication, many iterations, and a truly collective effort.
See you at SBBD!
In June, we had the opportunity to visit the Barcelona Supercomputing Center (BSC), where we were warmly hosted by Dr. Débora Pina.
The visit provided an excellent opportunity to learn more about the BSC and to visit MareNostrum 5, a pre-exascale EuroHPC supercomputer hosted at BSC, as well as the Quantum computer.
We sincerely thank Dr. Debora Pina for her warm hospitality.
As part of our internationalization initiatives and the activities of the DECoHPC Équipe Associée, members of the UFFeScience Research Group had the opportunity to visit INRIA Bordeaux and meet with Dr. Luan Teylo, a researcher from the TADaaM Group.
During the visit, Prof. Daniel de Oliveira and Wesley Ferreira, a Master's student at Universidade Federal Fluminense (UFF), discussed ongoing research on containerized workflow scheduling, particularly in the context of Wesley's Master's dissertation.
The meeting provided an excellent opportunity to exchange ideas, strengthen collaborations, and explore new research directions in scientific workflows across the computing continuum.
The visit also included two research presentations:
🔹 Prof. Daniel de Oliveira presented:
"Federated Outlier Detection for LSST-Scale Astronomical Data in Clouds"
The presentation discussed the challenges of applying Federated Learning to outlier detection in large-scale astronomical surveys, investigating performance, scalability, and cost trade-offs in distributed cloud environments.
🔹 Wesley Ferreira presented:
"AkôFlow: A Container-based and Provenance-Aware Workflow Engine for Portable Execution Across the Computing Continuum"
The talk introduced AkôFlow, an open-source workflow engine designed to support portable, provenance-aware, and resource-efficient execution of containerized workflows across heterogeneous environments, including HPC systems, clusters, and cloud infrastructures.
We thank the TADaaM Group and INRIA Bordeaux for the warm welcome and stimulating discussions. Initiatives such as this reinforce the importance of international scientific cooperation in advancing research and fostering innovation.
#UFFeScience #Internationalization #ResearchCollaboration #ScientificWorkflows #HPC #FederatedLearning #INRIA #TADaaM #DECoHPC
🚀 New Paper Accepted at SBBD 2026! 🚀
We are happy to share that our latest research paper, "Provenance-Guided Analysis of Agentic Workflows" (Análise de Workflows Agênticos Guiada por Proveniência -- In Portuguese), has been accepted for presentation at the Brazilian Symposium on Databases (SBBD 2026)! 🇧🇷✨
This work is a core part of Valesca Moura's M.Sc. dissertation at Instituto de Computação UFF under the supervision of Prof. Daniel de Oliveira (UFF) and Sandro Rama Fiorini (PUC-Behring Institute for Artificial Intelligence)
As Autonomous AI Agents and Large Language Models (LLMs) increasingly drive complex decision-making and automate workflows, a critical challenge emerges: How do we audit their choices and mitigate hallucinations when their reasoning remains hidden?
The proposed approach, ANCHOR (ANalysis of agentiC workflows via History and prOvenance), introduces a framework to solve this:
🎛️ Telemetry & Interaction-Centric Capture: Tracing execution spans down to individual agent intentions and actions.
📐 Data Structuring: Transforming unstructured logs into clear, hierarchical agent records (Interactions, Entities, Intentions).
🌐 W3C PROV Alignment: Standardizing data to export rich, graph-based provenance models for thorough workflow auditing.
Feel free to reach out at SBBD 2026 if you'd like to discuss agentic frameworks, data provenance, or AI auditability!
#SBBD2026 #AIAgents #LLMs #DataProvenance #DataEngineering #ResponsibleAI #UFF #Research #MScDissertation
🎉 Paper Accepted at SBBD 2026! 🎉
We are happy to share that our paper, "Personalized Service Level Agreements for Machine Learning Training using k-Skyband Queries", has been accepted at SBBD 2026 (Brazilian Symposium on Databases).
Training Machine Learning models in cloud environments involves important trade-offs between execution time, financial cost, energy consumption, and carbon emissions. Choosing the right infrastructure and resource configuration is often challenging, especially for large-scale ML workloads.
In this paper, we introduce an approach that extends the concept of Personalized Service Level Agreements (PSLAs) to Machine Learning training workflows.
By using provenance data and k-Skyband queries, our method automatically identifies resource configurations that represent different trade-offs between performance, cost, and sustainability, helping users make more informed decisions before launching training jobs.
This research was conducted as part of Gustavo Alexandre PhD at the Federal Fluminense University (UFF), within the UFFeScience Research Group. Gustavo is advised by Prof. Daniel de Oliveira and Prof. Yuri Frota. We also thank Prof. Marcos Bedo for his valuable collaboration and contributions to the paper.
Looking forward to presenting our results at SBBD 2026 and discussing future directions for sustainable and data-driven Machine Learning infrastructure management.
#SBBD2026 #MachineLearning #CloudComputing #Databases #eScience #Provenance #Sustainability #ArtificialIntelligence #Research #PhD
We are excited to share that Valesca Moura, a Master's student at the Institute of Computing@UFF (Instituto de Computação UFF) and a member of UFFeScience Research Group, has started an internship at the Oak Ridge National Laboratory (ORNL)!
During her internship, Valesca will work on the topic of her Master's dissertation, "Provenance Management for Agentic Workflows", collaborating closely with the ORNL research team to advance this important area of study.
Her work will be supervised by Renan Souza, strengthening the collaboration between IC/UFF, UFFeScience, and ORNL.
This internship offers an excellent opportunity for international research collaboration and contributes to advancing research on agentic systems and provenance management.
Congratulations, Valesca, on this achievement! We wish you a successful and rewarding experience at ORNL. 🎉
#UFF #UFFeScience #ORNL #Research #Internship #AgenticAI #ProvenanceManagement #ScientificCollaboration
🇧🇷 Como grupo de pesquisa, temos grande satisfação ao compartilhar a defesa de mestrado de nosso integrante, Thiago Jordão, membro do UFFeScience Research Group. Seu trabalho, intitulado "ARANI: Uma Abordagem Baseada em Linha de Experimento para Preservação de Privacidade e Auditoria em Data Lakes", reforça o compromisso do grupo com pesquisas relevantes e de impacto na área de dados.
A banca examinadora foi composta pelos professores Javam Machado (UFC) e Vanessa Braganholo (IC/UFF). O trabalho foi orientado pelos professores Daniel de Oliveira e Marcos Bedo.
Parabéns, Thiago, por essa importante conquista! 👏🚀 Temos orgulho de contar com você no UFFeScience Research Group e desejamos muito sucesso nos próximos passos da sua trajetória.
🇺🇸 As a research group, we are proud to share the Master's defense of our member Thiago Jordão, part of UFFeScience Research Group. His work, entitled "ARANI: An Experiment Line-Based Approach for Privacy Preservation and Auditing in Data Lakes", highlights our group’s commitment to impactful research in the data domain.
The examination committee was composed of Professors Javam Machado (UFC) and Vanessa Braganholo (IC/UFF). The work was supervised by Professors Daniel de Oliveira and Marcos Bedo.
Congratulations, Thiago, on this important achievement! 👏🚀 We are proud to have you at UFFeScience Research Group and wish you continued success in your journey.
🚀 E se fosse possível executar workflows científicos em qualquer ambiente, sem dor de cabeça com a compatibilidade e a portabilidade?
Apresentamos o AkôFlow: um novo jeito de gerenciar workflows científicos em qualquer plataforma.
Temos o prazer de compartilhar um projeto em desenvolvimento que já demonstra grande potencial de impacto nas áreas de computação científica e de engenharia de dados.
🎓 O trabalho faz parte da dissertação do Wesley Ferreira, sob orientação do Prof. Daniel de Oliveira e de colaboradores: Profa. Aline Paes, Prof. Marcos Bedo, Profa. Liliane Kunstmann, Prof. Yuri Frota, Luan Teylo, Raphael Reis Garcia e Rodrigo A. P. da Silva.
O AkôFlow nasce para resolver um desafio bem conhecido por quem trabalha com workflows complexos:
👉 Como executar workflows em diferentes ambientes (clusters, nuvem, HPC) sem sofrer com incompatibilidades e dependências?
💡 Nossa proposta combina conteinerização + escalonamento para entregar:
✔️ Portabilidade entre ambientes heterogêneos de forma transparente
✔️ Execução compatível com Docker, Kubernetes e Singularity
✔️ Reutilização de workflows existentes
✔️ Monitoramento completo com dados de proveniência
Na prática, o mesmo workflow pode ser executado em diferentes infraestruturas: sem retrabalho!
📌 Mais do que rodar tarefas, estamos falando de garantir reprodutibilidade, eficiência e escalabilidade em experimentos científicos.
Assista às apresentações do AkôFlow:
🎥 https://lnkd.in/d9vrNzqS
🎥https://lnkd.in/dj_p6ixt
Os artigos associados ao AkôFlow podem ser obtidos em:
📚 https://lnkd.in/deUJ_XiF
📚 https://lnkd.in/d2XHUtdc
📚 https://lnkd.in/dptpQa7t
Mais informações em: https://akoflow.com/
Esse é só o começo! Seguimos evoluindo o AkôFlow com novas funcionalidades, integrações e contribuições open source.
Fica o convite: vamos repensar juntos como construímos e executamos workflows?
#Akoflow #computacaocientifica #cienciadedados #opensource #kubernetes #docker #pesquisa #inovacao
We are happy to share that our paper, “When Prompts Know the Story: Improving LLM Detection of Online Hate Speech,” has been accepted for publication in the Journal of Information and Data Management (JIDM) and will be published soon.
This work is the result of Laryssa Paiva's undergraduate thesis (TCC), and it is great to see this research reach the community through a journal publication.
We would like to thank all the co-authors (Laryssa Paiva, Annie Amorim, Gabriel Assis, Luiz Gustavo Dias, Aline Paes, and Daniel de Oliveira) for their collaboration and contributions throughout this work.
Research is always a collective effort, and this result reflects the dedication of everyone involved.
This paper is also the outcome of a collaboration between UFFeScience Research Group and MeLLL@UFF!
Looking forward to sharing the final version once it is officially published! 🚀
AkôFlow is an open-source engine for orchestrating and executing container-based scientific workflows in the computing continuum. It was originally developed within the e-Science Research Group at the Institute of Computing, Fluminense Federal University (UFF).
You are visitor 01402503 — GeezerVison does not store the pages you look up. Best viewed with Netscape Navigator 4.7 or Internet Explorer 6 at 800×600. No cookies required.
UFFeScience Research Group
