M.Sc. student in Computer Science (AI) at the Federal University of Minas Gerais (UFMG), Brazil, and research fellow at Kunumi. I work where HPC, LLM systems and machine learning meet: making many models share scarce GPUs, building and evaluating LLM pipelines and agents, and training models that need fewer labels and less compute.
Multi-LLM serving on shared GPUs (prism-research), my M.Sc. research. I extend and stress-test Prism, an open-source system that serves many LLMs on one or more GPUs by time-sharing or splitting memory (SGLang + kvcached). Experiments run on NVIDIA H200 GPUs through a Polyaxon cluster.
- vLLM engine backend with a persistent worker pool and tensor parallelism (TP=2): completed requests went from 62% to 100% on the 2-GPU example, with latency on par with SGLang.
- Model eviction policies evaluated over 57+ runs on replayed request traces: p90 TTFT 53 s → 0.6 s and SLO attainment 51% → 90%. A throughput-per-GB swap policy was the only setup that served 100% of requests in every scenario.
- Out-of-memory robustness: diagnosed five failure modes (crashed workers, OOM server crashes, full stalls) and fixed them on the Prism side, with dead-worker failover in ~4 s, OOM rollback and a cross-process memory arbiter. A memory-pressure scenario went from 70/337 requests with a server crash to 337/337.
Every experiment has a dated report with the scripts and commands to reproduce it.
| Project | What it is |
|---|---|
| prism-research | Research fork of Prism: vLLM backend, eviction policies and memory-robustness fixes for multi-LLM serving. |
| isadg-agents | Multi-agent pipeline (LangChain) with 4 specialized LLM/VLM agents (vision OCR, semantic router, normalizer, RAG archivist) that catalogs 19th-century documents under the ISAD(G) archival standard. 95–99% word-level OCR accuracy on 917 pages, with retries, self-healing and per-agent latency profiling. |
| VoxLab | Reproducible LLM evaluation of stance-change detection in Brazilian public hearings: a contrast set checked by LLM verifiers, 7 open-weight models × 5 prompting architectures, frozen prompts and a content-hashed response cache so the whole study replays in under a minute with no API calls. |
| seismic-fsss | PyTorch pipeline for episodic few-shot segmentation of 3D seismic volumes, with configurable data partitioning and support-selection strategies. |
| Seismic-Segmentation-Framework | TensorFlow benchmark that unifies training and evaluation of deep learning models for seismic facies segmentation. |
- G. L. Nunes et al. Spatial Data Partitioning and Support Locality in Few-Shot Seismic Facies Segmentation. First author, 2026 (submitted).
- G. Amarante, M. T. P. Souza, G. Lima et al. Enhancing Minority Class Recovery in High Resolution Land Cover Mapping with Dynamic Imbalance Aware Oversampling. Scientific Reports, 2026. doi:10.1038/s41598-026-52071-z
- L. M. Laboissiere, G. L. Nunes, C. T. Cunha. Comparability Between Statements of the Same Actor: A Contrast Set for Stance-Change Detection with LLMs. 1st Kunumi Data Challenge, 2026.
- G. Lima et al. A Systematic Evaluation Methodology of Deep Learning on Seismic Facies Classification. First author, Earth Science Informatics, 2025. doi:10.1007/s12145-025-02024-3
- LLM systems: vLLM, SGLang, KV-cache management, tensor parallelism, LangChain, RAG (ChromaDB)
- ML: Python, PyTorch, TensorFlow, sentence-transformers
- HPC & infrastructure: CUDA GPUs, Polyaxon, Docker, Kafka, Linux
I'm interested in work on LLM inference and serving, HPC and ML systems. Email: [email protected]