Wyctor Fogos da Rocha
Projects

The work behind the papers

Each case study below walks through the problem, the architecture, the evaluation and the results. For the complete list of publications with abstracts and BibTeX, see publications.

Case studies

Featured projects

  • IJCNN 2026 First author

    RG-DermNet

    Residual gated attention for multimodal skin lesion classification

    A fusion block that lets clinical metadata modulate the visual features without ever cutting off the image pathway — so the model keeps working when the anamnesis is incomplete. AUC 0.95 ± 0.01 on PAD-UFES-20.

    • Multimodal learning
    • Attention
    • Interpretability
    Read the case study
  • Scientific Reports Co-author

    PRISM

    Probabilistic reasoning for interpretable diagnosis

    A Bayesian framework that lets a clinician watch the diagnosis shift as each piece of evidence arrives, with a stepwise calibration protocol that keeps confidence honest. Peak balanced accuracy 77.2 ± 3.6 %.

    • Bayesian inference
    • Calibration
    • Explainable AI
    Read the case study
  • IEEE JBHI Co-author

    MetaBlock-SE

    Sentence embeddings for missing clinical metadata

    Replacing MetaBlock's sparse categorical encoding with a sentence embedding, so the model captures how clinical attributes relate — and survives when some are absent. Better than the original in every scenario tested.

    • Missing data
    • Sentence embeddings
    • Efficiency
    Read the case study
  • ICECCME 2025 First author

    SRGAN-upscaled YOLO

    Super-resolution as a training strategy for aerial detection

    Aerial scenes mix very large and very small objects at uneven image quality. Upscaling the weakest images with an SRGAN during training raises detection performance across ten YOLO architectures on DOTA v1.5.

    • Object detection
    • Super-resolution
    • Benchmark
    Read the case study
In progress

Ongoing work

Research still in progress, with the code public while it runs. The results are not settled yet — treat these as work you can read, not as conclusions.

  • MSc project · PPGI/Ufes

    LLM as a NAS controller

    Neural architecture search for multimodal skin lesion classification, driven by a local LLM served through Ollama. At each step the model receives the search space and the history of what has already been tried, proposes a new architecture configuration as JSON, and gets the resulting balanced accuracy back as the reward that shapes its next proposal.

    llm-as-nas-controller
  • Paper in preparation

    NAS optimization for a multimodal VLM

    The experimental code behind Neural Architecture Search optimization for a multimodal model for skin lesion recognition — a case study: searching over vision-language configurations on PAD-UFES-20 under patient-wise 5-fold cross-validation.

    NAS-optimization-for-VLM
Open source

Code on GitHub

A selection from my repositories — the systems and experiments I keep public, grouped by what they do.

Search & architecture optimisation

  • NAS-ML-optimization

    Architecture search with Microsoft NNI on CIFAR — the groundwork for the LLM-driven controller.

    Python
  • fine-tune-tokenizer

    Adapting a tokenizer to a domain vocabulary before fine-tuning the model on top of it.

    Python
  • simple-image-classifier-pytorch

    A minimal, readable PyTorch training loop — the baseline I start experiments from.

    Python

LLM services & document intelligence

  • AUTOMATIC-OCR-SERVICE

    A Docker Compose OCR service that watches an input folder and emits the extracted content as JSON.

    Python
  • RAG-DOCS-with-PGVECTOR

    Retrieval-augmented generation over documents, with embeddings stored in Postgres via pgvector.

    Python
  • RESUME-INTELLIGENTE-ANALYZER

    Job recommendation by RAG: sentence-transformer embeddings indexed in FAISS, answers generated with Ollama.

    Jupyter
  • NER

    End-to-end named entity recognition: dataset generation, training and inference scripts.

    Python
  • agent-ai-with-langchain

    A LangChain agent running against a self-hosted Ollama model, configured entirely through env vars.

    Python
  • ia-transc2form-with-ollama

    Turning a transcription into a filled form, with a local Ollama server exposed to the network.

    JavaScript

Fine-tuning small models

Computer vision systems

All repositories on GitHub

Also published

Other work

Published research without a case study page yet — the abstracts and BibTeX are on the publications page.

  • Verbose metadata descriptions generated by LLMs

    LLMs as semantic transducers: structured clinical attributes rewritten as anamnesis-like text, encoded with SBERT and fused through MetaBlock-SE. SBCAS 2026

  • LiwTERM-r

    A lightweight transformer for multimodal skin lesion detection, robust to incomplete clinical input and to multiple images of the same lesion. Journal of the Brazilian Computer Society, 2026

  • Virtual reality diagnostic room for lower limb amputees

    A multi-camera intelligent space feeding a VR environment where physiotherapists inspect gait parameters, with a neural classifier averaging above 91 % across evaluation metrics. Artificial Intelligence in Medicine, 2023

  • Gait recognition with the StarRGB technique

    Condensing the spatial-temporal information of a gait sequence into a single image so a CNN can classify movements in remote physiotherapy. ICECCME 2021

  • Anomaly detection in civil structures

    A benchmark of machine learning models for detecting anomalies in structural monitoring data. ICMLA 2023

  • Books on systems analysis, production planning and inventory

    Three teaching books published with Atena Editora, used as course support material in technical and undergraduate programmes. Atena Editora, 2023–2024

Every paper, with abstracts and BibTeX

Fifteen entries across multimodal learning, medical AI, computer vision and embedded systems.

View publications