Fastlane is a Retrieval-Augmented Generation (RAG) system designed as an intelligent assistant for women’s clothing stores in Brazil. It recommends products to customers, facilitates purchases, and processes payments quickly and securely—all within a familiar platform like WhatsApp. The assistant enhances the shopping experience by leveraging audio, images, and text. On the technical side, Fastlane is being built with Python for backend services, the WhatsApp API for seamless app integration, large language models for generative AI, and Elasticsearch as a vector database to enable semantic product search.

Fastlane is being collaboratively built with my friend and colleague from PUC-Rio, Leonardo Robinson and you can find more information here.

Technology stack: Python Elasticsearch Large Language Models Whatsapp Cloud API Natural Language Processing

Smart Tocha is a computer vision solution that aims to increase the efficiency of burning gases in the torch of refineries. Applying machine learning techniques, Smart Tocha processes burning images continuously, classifies the current state of the torch and acts directly on the industrial control system. The system monitors and controls the steam flow in real time, unlike manual control. This artificial intelligence keeps burning safe for the environment, minimizing greenhouse gas emissions.

The solution caught attention in various media reports:

• Petrobras (in Portuguese)
• DI PUC-Rio (in Portuguese)

The Smart Tocha project was developed through a collaborative effort between the ExACTa PUC-Rio initiative and the research center of Petrobras, a leading Brazilian oil company. The source code is not provided due to intellectual property restrictions.

Technology stack: Python Azure Machine Learning Workspace Jupyter Notebook Stream Analytics Azure Functions PyTorch TensorFlow OpenCV

Artificial intelligence techniques, particularly Convolutional Neural Networks, have found widespread application in addressing various computational vision challenges. Motivated by the potential of neural networks to discern patterns in images, I endeavored to create a deep learning model aimed at identifying the authors of renowned paintings.

As this project is a personal endeavor aimed at learning and experimentation, I utilized Gradio and HuggingFace to develop and distribute this application. Gradio provided an intuitive interface for interaction, while HuggingFace facilitated access to pre-trained models and streamlined the development process. Together, these tools enabled me to create an efficient and user-friendly application for exploring and utilizing deep learning techniques in the context of painting authorship.

The project's source code is available in the GitHub repository.

If you wish to access the deployed app, please visit the HuggingFace space where the project is hosted.

Technology stack: Python Convolutional Neural Networks Gradio Jupyter Notebook Object-oriented programming HuggingFace

IAGO (Artificial Intelligence for Odor Management) is a machine learning project designed to take a proactive approach by alerting the refinery's operations and engineering team about the increased likelihood of emissions that may lead to community discomfort. By identifying patterns in operational variables and meteorological conditions, IAGO offers insights to assist refinery technicians in swiftly diagnosing issues and taking preventative measures to mitigate odor nuisances.

The Iago project was developed through a collaborative effort between the ExACTa PUC-Rio initiative and the research center of Petrobras, a leading Brazilian oil company. The source code is not provided due to intellectual property restrictions.

Technology stack: Python PowerBI Azure Machine Learning Workspace Jupyter Notebook Stream Analytics Azure Functions Logistic Regression

I aimed to deepen my understanding of applying deep learning to both binary and multiple image classification tasks. To achieve this, I developed two Jupyter notebooks: one for binary classification using images of cats and dogs, and another for multiple classification involving images of household objects. These notebooks include features to augment images, train models, and evaluate their performance.

The project's source code is available in the GitHub repository.

Technology stack: Convulational Neural Network Jupyter Notebook Data Augmentation