AI / ML Engineer — from research to production.
I build LLM-based systems and multimodal pipelines that run in production, with end-to-end MLOps: fine-tuning, containerised serving, monitoring. My peer-reviewed research (incl. Scientific Reports, IEEE JBHI) grounds the engineering in statistical rigour.
Currently at Tribunal de Justiça do Espírito Santo (TJES), where I built LuzIA — an LLM-powered document analysis system that cut ruling review from 10 to 2 minutes and was selected by Brazil's National Council of Justice (CNJ) for nationwide rollout¹.
🌐 wyctor.com.br · LinkedIn · Google Scholar · ORCID
- LLM systems for the judiciary — multi-stage NLP pipelines (extraction, cross-encoder re-ranking, hybrid scoring) for legal document classification, served with FastAPI + RabbitMQ + vLLM + Docker.
- Multimodal medical AI — gated cross-attention architectures for skin lesion classification, published at IJCNN 2026 (IEEE WCCI) and in Scientific Reports (Nature Portfolio).
- LLM-driven neural architecture search — using local LLMs as NAS controllers under deployment constraints (SIBGRAPI 2026).
| Project | What it does | Links |
|---|---|---|
| anon-lib | PII detection & masking for pt-BR and English — hybrid regex + local LLM, 21 pt-BR / 7 EN entity types, first library to support the 2026 alphanumeric CNPJ. | PyPI · GitHub |
| RAG-DOCS-with-PGVECTOR | Retrieval-augmented generation over documents, embeddings stored in Postgres via pgvector. | GitHub |
| RG-DermNet | Residual gated-attention multimodal fusion for skin lesion classification — AUC 0.95 ± 0.01 on PAD-UFES-20. | Paper (IJCNN 2026) |
| LLM-as-NAS-controller | Neural architecture search driven by a local LLM served through Ollama. | GitHub |
| Fine-tuning Gemma 3 1B | LoRA fine-tuning with Unsloth, exported for vLLM inference. | GitHub |
| NER | Named entity recognition with a BERT-based model for text anonymisation. | GitHub |
More repos: github.com/wyctorfogos?tab=repositories
- PRISM — clinically interpretable stepwise framework for multimodal skin cancer diagnosis. Scientific Reports (Nature Portfolio), 2026. DOI
- RG-DermNet — multimodal attention-based model for skin lesion classification. IJCNN 2026 (IEEE WCCI). Preprint
- MetaBlock-SE — handling missing metadata via sentence embeddings. IEEE JBHI, 2025. DOI
- LLM-driven NAS — architecture search under deployment constraints. SIBGRAPI 2026. GitHub
Full list: wyctor.com.br/publications
Languages: Python (production-level), SQL, JavaScript
ML & LLM: PyTorch, Hugging Face Transformers, LLM fine-tuning (LoRA/PEFT), RAG, cross-encoders, vLLM, Ollama, OpenCV, MLflow
Backend & infra: FastAPI, Flask, gRPC, Docker, RabbitMQ, Kafka, PostgreSQL, Azure, Git
Data: PySpark, Pandas, NumPy
- M.Sc., Informatics — Data Science and AI · Federal University of Espírito Santo (UFES), 2026
- Télécom SudParis, France · M2 TRIED coursework, BRAFITEC exchange, 2021–2022 (63 ECTS)
- B.Sc., Electrical Engineering · Federal Institute of Espírito Santo (IFES), 2023
📫 wyctor.fogos@gmail.com · Based in Vitória, Brazil · Trilingual: Portuguese · English · French