I'm a Software Engineer and an Engineering Physics graduate from the University of Campinas (UNICAMP), currently pursuing a degree in Computer Engineering.
My work focuses on the intersection of software engineering, artificial intelligence, and applied research. I'm particularly interested in building reliable AI systems, turning machine learning research into working solutions, and solving complex real-world problems.
I have professional experience developing optimization software and AI-powered applications, alongside academic research in computer vision, reinforcement learning, and natural language processing.
I work on software solutions for logistics optimization, decision support, and artificial intelligence.
Some of my main contributions include:
- RAG & LLM Systems: Designed the architecture and led a four-developer team building an enterprise Retrieval-Augmented Generation platform. The system integrates hybrid retrieval, LLM-based routing, document processing, and answer grounding. It is operational in a controlled internal pilot.
- Cloud & AI Infrastructure: Worked with Azure Functions, Cosmos DB, Key Vault, and Application Insights to support AI services.
- Software Architecture: Develop and maintain C#/.NET applications and contribute to architectural modernization initiatives.
- Optimization: Improve algorithms and heuristics for logistics planning and operational decision-making.
- AI-Assisted Development: Introduced and promoted AI-assisted software development practices within the company, including developer training, repository guidelines, reusable agent skills, and safe development workflows.
My research explores visual perception and learning-based robotic manipulation.
I have worked on adapting the VAPO framework with DINOv3-based visual representations, developing evaluation tools and conducting reproducible experiments to investigate the relationship between visual perception quality and downstream robotic performance.
Research interests: Computer Vision, Vision Transformers, Reinforcement Learning, Robotic Manipulation, Representation Learning, and Experimental Evaluation.
My NLP work began with a collaborative academic project at UNICAMP on information extraction from clinical case reports and knowledge graph construction, which I later extended individually.
The work includes classical NLP techniques, structured information extraction, entity and relation identification, and comparative evaluation of machine learning approaches.
Undergraduate research at UNICAMP integrating DINOv3 into VAPO's gripper-camera affordance perception pipeline for robotic manipulation. My fork builds on the original VAPO authors' work, adding encoder integration, evaluation tools, and compatibility tests.
- Perception: Evaluation of 6,386 paired images yielded global binary mIoU of 0.725 for the baseline and 0.819 for DINOv3.
- Robotic performance: Across 3,000 simulated episodes (1,500 per approach), success was 86.20% for the baseline and 62.47% for DINOv3. Better segmentation did not translate into better task performance in these experiments.
- Analysis and reproducibility: Investigated observation distribution shifts, compatibility with the existing Soft Actor-Critic (SAC) policy, and manipulation-point localization as possible explanations; these hypotheses remain unconfirmed. Documented experiments, automated checks, artifact hashes, and reproduction limits.
Comparison limits: Input configurations, resolution, and label generation also differ between approaches; these results do not isolate the encoder's effect.
Technologies: Python, PyTorch, DINOv3, Computer Vision, Deep Learning, Affordance Learning, Reinforcement Learning (SAC).
UNICAMP academic project for extracting entities and relations from clinical case reports and constructing knowledge graphs. The original project and core NLP pipelines were developed with Matheus Rufino da Silva, Daniel de Sousa Cipriano, and Eduardo dos Santos Neres.
- Comparative evaluation: Four pipelines combine dictionary or CRF entity extraction with rule-based or k-NN relation extraction, evaluated through supervised metrics, cross-validation, and error analysis.
- My individual extensions: Built a local Streamlit demo and a deterministic synthetic example; reran and consolidated five-fold nested cross-validation; audited experimental artifacts and expanded reproducibility documentation. The repository includes automated tests and continuous integration.
- Interactive demo: Live entity extraction, relation extraction, and graph construction use the classical dictionary-and-rules pipeline. CRF and k-NN results are displayed from recorded evaluations, without live supervised-model inference.
Technologies: Python, Streamlit, Conditional Random Fields (CRF), k-Nearest Neighbors (k-NN), NLP, Knowledge Graphs, Automated Testing.
Collaborative property rental application developed for UNICAMP's MC322 course, with a Java backend and React frontend. The project covers property, user-profile, and rental management, applying object-oriented design and full-stack development.
Technologies: Java, React, JavaScript, React Router, Context API, CSS.
Programming: Python, C#, Java, SQL, JavaScript
Machine Learning & AI: PyTorch, Computer Vision, Reinforcement Learning, Natural Language Processing, LLMs, RAG, Embeddings, Vector Search, BM25
Software Engineering: .NET, REST APIs, Software Architecture, Automated Testing, Git
Cloud & Databases: Azure Functions, Azure Cosmos DB, Azure Key Vault, Application Insights, SQLite
Additional Technologies: React, Flutter, Linux
- Engineering Physics — University of Campinas (UNICAMP), graduated in 2026
- Computer Engineering — University of Campinas (UNICAMP), currently enrolled
I enjoy working on problems that require both analytical thinking and practical engineering.
My goal is to keep growing as an AI Engineer, combining strong software engineering foundations with machine learning research to build intelligent systems that are reliable, measurable, and useful in the real world.