Felipe Costa Farias

1. A natural language processing (NLP) system written in Python using state-of-the-art Machine Learning algorithms, with Scikit-Learn. The model helps customer service teams by suggesting the top three most plausible responses to questions/emails sent by clients. The model trains on previously answered emails written by each team. Analysts were able to increase their response time by 12x. The selection of ideal responses occurs in 30% of the cases. This number increases while more data are gathered and presented during retraining. This project uses a lightweight Flask web service and works for English, Portuguese, and Spanish languages.

2. A system that performs sentiment analysis using state-of-the-art machine learning algorithms on several social media texts with SpaCy/Gensim. It currently handles thousands of requests per second using a lightweight Flask web service.

3. Module to identify a person (i) age and (ii) gender using social media profile photos applying deep convolutional neural networks with PyTorch.

4. Module for automatic Tagging of social media posts based on previously tagged ones written in Python with scikit-learn.

A SaaS to perform classification of Electrocardiogram (ECG) data. The ECG data was acquired by custom hardware attached to a smartphone. The smartphone communicates with a web service endpoint to send the data to further AI analysis. I was the AI consultant leading an AI team to create and deploy the models.

In this project, I was a software consultant for robotics, mentoring many engineering interns from all over the USA and several other countries. It was a project where we had to develop the hardware/software of robots to collaboratively create 3D maps. Each robot has a LIDAR camera that takes 360° point cloud pictures of its location. Each robot has two spheres with unique distances on top of them, serving as a visual identifier that we had to detect in the point cloud. As the robots took several images from different overlapping areas, we performed image registration using SLAM, taking as landmarks the spheres position, fusing data from these landmarks and each robot motor encoder, as we have no access to GPS data. After this, we had to set the next best location to send each robot to take other pictures, maximizing the area over time. To perform the path planning, we have used a classical AI algorithm (A*). We implemented the software using Robotic Operating System (ROS) and OpenCV in C/C++ and Python. We also used the Point Cloud Library. To filter the point cloud data, we have used signal processing techniques. To visually analyze our data, we have used the rviz package and custom OpenGL software.

A system to classify imagined and executed movements using deep learning algorithms such as convolutional neural networks, long short-term memory, and gated recurrent units from subjects' electroencephalography (EEG) input signals. It was built in Python with the Keras framework.

A time-tracking system developed for lawyers' activities and support a law office's operations, controlling all the documents with several reports built to support the work audit.

We used Django and PostgreSQL to create both systems.

A flexible ML framework that enables non-ML expert people to train their models based on their own data and facilitate the implementation of new training tasks and model architecture definitions for the ML team regarding molecular property prediction. We also coupled this system with automated training such that the models are always up-to-date.