Machine Learning Developer in Florianópolis - State of Santa Catarina, Brazil

While working on a Toptal project for Bractlet (an award-winning company focused on modeling building energy usage in order to improve energy efficiency), I developed an application which uses machine learning to calibrate physics-based energy models.

These models are very powerful, but they take a long time to run and contain a number of parameters which must be calibrated and are not known ahead of time. Thus the objective of the application was to automate the calibration of these parameters using as few iterations of the physics-based model as possible.

For this project, I developed an application that uses machine learning to model the parameter space and select parameter sets to use in simulations, simultaneously exploring the parameter space and minimizing the physics-based model error without human input.

I worked with a team at Bravi to build a predictive model for a Brazilian university in order to indicate those students at risk of dropping out.

We cleaned and extracted features from the raw university data, evaluated the performance of various machine learning algorithms and the models they produced, and incorporated the resulting models into a Docker image which is currently in use at the university. The predictions are then used to focus early attention on students who are at risk of dropping out and maximize their chance of continuing their studies.

My first experience with machine learning was at the end of my undergrad time at Princeton when a friend approached me to help him implement a model to predict the outcome of baseball games, which he had designed as part of his senior thesis.

We used non-parametric statistics (before machine learning was a buzzword), and custom built a model to predict the probabilities of certain events occurring in a particular game. The model was a nearest neighbor's implementation using composite indices for dimensional reduction.

Treating baseball betting as a market, we used the model to trade very successfully for two years before new laws made the market unavailable.

At Way2, I worked with an Agile team to build a platform to collect data from energy meters throughout Brazil and South America.

We designed and built a scalable microservice solution which is highly parallel and asynchronous, robust for longterm stability, can communicate using a variety of communication protocols, and has detailed logging.

This platform is currently in use by CCEE, the Brazilian government agency which manages the Brazilian energy market, collecting data from tens of thousands of meters.

The purpose of this repository is to show how I like to develop data science projects and what can be built in a few hours. It includes two Jupyter notebooks—the first showing exploratory analytics and discussion of the data, and the second showing the performance of predictive models built via machine learning.

The performance of some standard machine learning algorithms are compared to that of a custom-designed model tailored to the system being modeled. The custom model results in a reduction of nearly half the residual error between the prediction and the testing data.

Verbalist was a productivity app for Android available on the Google Play App Store.

It was a list manager with voice-to-text and semi-structured language processing which allowed users to control the app and add list items by voice. The app had thousands of users and a 4.8 star rating until we stopped development.