Experience
Di2win - R&D Engineer
Dec de 2021 - Now
I developed methods to improve the energy efficiency of production lines in the industry. My primary responsibilities involved developing prediction and prescription models to recommend production line parameters, ensuring both quality output and operational efficiency for the industry. I work with models and methods such as Clustering, Optimization and Transformers.
Bosch Portugal & University of Minho - Research Assistant
May de 2020 - Dec de 2021
I developed a real-time activity recognition pipeline for efficient video classification. My responsibilities focused on improving accuracy by combining different video features — like RGB frames, audio, and optical flow. Specifically, I worked with models such as C2D, I3D, SlowFast, X3D and developed custom ones to integrate the different video features.
Publications

Exploring Image Classification Robustness and Interpretability with Right for the Right Reasons Data Augmentation
LXAI Workshop at ICCV 2023
Evaluating zero-shot image classification based on visual language model with relation to background shift
LXAI Workshop at NeurIPS 2023
Towards Background and Foreground Color Robustness with Adversarial Right for the Right Reasons
ICANN 2023
Image Classification Understanding with Model Inspector Tool
HAIS 2023
On the Impact of Interpretability Methods in Active Image Augmentation Method
Logic Journal of IGPL 2022
A hybrid post hoc interpretability approach for deep neural networks
HAIS 2021Projects
Model inspector tool
This tool allows users to load image classification models from Pytorch and Timm library, then they can manipulate various visual features of an input image to understand better the model’s sensitivity to different types of information. Our goal is to provide a more comprehensive framework for model understanding and help researchers and practitioners better understand the strengths and weaknesses of deep learning models in image classification.