Rodrigo Lima
- rsl@cin.ufpe.br
- rodrigoliimaa.github.io
- 0000-0002-8103-7723
- Recife/Maceió · Brazil
PhD Candidate in Computer Science at the Federal University of Pernambuco (UFPE), advised by Prof. Leopoldo Texeira and Co-advised by Prof. Baldoino Fonseca. My research investigates harmful code, the temporal relationship between code smells and bugs, and how Large Language Models change software development and review practices.
Recent and ongoing work covers (i) longitudinal analysis of how smells precede bugs in OSS projects, (ii) the use of small open LLMs for detecting defect-prone code smells, and (iii) how the adoption of LLM-assisted development reshapes pull-request review regimes. I also collaborate on applied research with NTS-AL / EASY-UFAL on telehealth platforms for high-risk obstetric care (project GESTAR).
Research Interests
- Empirical software engineering: longitudinal studies, mining software repositories
- Code smells, harmful code, and the temporal evolution of software defects
- LLMs for software engineering: detection, generation, and review
- Pull-request review regimes in the era of LLM-assisted development
- Applied research in healthcare informatics (telehealth, clinical decision support)
Publications
Can Small LLMs Detect Defect-Prone Code Smells?
Empirical study evaluating four open small LLMs (Qwen3-Coder-30B, DeepSeek-V2-16B, Llama-3.1-8B, Granite-3.1-8B) for detecting defect-prone code smells, with trivial baselines, null-baseline permutation tests, and determinism tests under quantized inference. Results show that small LLMs match or exceed simple rule baselines while remaining cheap to run locally — with caveats around determinism and prompt sensitivity.
R. Lima, J. Souza, B. Fonseca, L. Teixeira, M. Ribeiro. Artifact (Zenodo)
GESTAR: An Integrated Telehealth Platform for Obstetric Risk Stratification, e-Learning, and Health Information Management
An integrated telehealth platform combining rule-based obstetric risk stratification (73 factors from the SESAU/AL NT15/2025 protocol), continuing education for primary-care professionals, and clinical decision support, deployed across seven municipalities in Alagoas. 354 risk stratifications recorded between Oct/2025 and Feb/2026; 33.9% classified as High Risk, 25.7% Medium, 40.4% Usual.
R. Lima, T. L. Ferreira, M. C. Oliveira, D. M. Baia, P. Pimentel, B. Fonseca, M. Ribeiro
GESTAR: A Telehealth Platform for High-Risk Pregnancy Triage and Longitudinal Follow-up
Tools-track companion paper describing GESTAR's microservices architecture (Python/FastAPI, PostgreSQL, Next.js, Docker), security model (AES-256, HMAC-SHA256 blind indexes, role-based access control under LGPD), and deployment lessons learned from rural-municipality rollout in Alagoas.
R. Lima, T. L. Ferreira, M. C. Oliveira, D. M. Baia, P. Pimentel, B. Fonseca, M. Ribeiro
Exploring Transfer Learning for Multilingual Software Quality: Code Smells, Bugs, and Harmful Code
Empirical investigation of cross-language transfer learning for software-quality prediction over 641K+ commits across Java, C++, C#, and Python. The study evaluates how models trained on one language and one outcome (bugs or smells) generalize to another, finding transfer between bugs and code smells to be a not-ineffective avenue for detecting harmful code in low-resource languages.
R. Lima, D. Pereira, C. Barbosa, L. Leite, D. Baia, B. Fonseca, L. Teixeira, J. Souza · 10.5753/jserd.2025.4593
Investigating the Social Representations of Harmful Code
Qualitative study of how software developers perceive harmful code, using free-association tasks with Brazilian postgraduate students and industry practitioners. The investigated community strongly associates harmful code with a core set of undesirable characteristics — bugs and code smells of various kinds — informing how future tooling and research should frame the construct.
R. Lima, J. Souza, B. Fonseca, L. Teixeira, R. Mello, M. Ribeiro, R. Gheyi, A. Garcia · 10.5753/jserd.2024.3554
Do You See Any Problem? On the Developers' Perceptions in Test Smells Detection
Empirical study on how developers perceive 10 types of test smells, with 25 participants performing over 1,250 assessments. Developers exhibited low agreement when detecting test smells, with specific heuristics — rather than experience level — driving their perceptions, suggesting that automated detection tools should account for this human-perception variability.
R. Lima, K. Costa, J. Souza, L. Teixeira, B. Fonseca, M. D'Amorim, M. Ribeiro, B. Miranda · SBQS 2023
Developers' Viewpoints to Avoid Bug-Introducing Changes
Mixed-methods study applying Q-methodology with 41 developers and 41 assumptions extracted from literature and personal interviews, to investigate which assumptions developers actually rely on to avoid bug-introducing changes. Five distinct viewpoints emerged, alongside a shared consensus on the importance of being aware of changes that invoke a large number of features.
J. Souza, R. Lima, B. Fonseca, B. Cartaxo, M. Ribeiro, G. Pinto, R. Gheyi, A. Garcia · 10.1016/j.infsof.2021.106766
Understanding and Detecting Harmful Code
Empirical study introducing the notion of harmful code: code already linked to bugs and still prone to harm software quality. Over 803 versions of 12 OSS projects, 132,219 smells and 40,340 bugs, only 0.07% of smells were harmful. Random Forest classifiers achieved >97% effectiveness in classifying harmful code; both software and developer metrics matter.
R. Lima, J. Souza, B. Fonseca, L. Teixeira, R. Mello, M. Ribeiro, R. Gheyi, A. Garcia · 10.1145/3422392.3422420