Data Scientist and Machine Learning Developer in Toronto, ON, Canada

Trained generative adversarial networks (GAN) using a dataset of celebrities' photos to learn how to create a fake human face. This same technique can generate real content like images, texts, simulations, or signals.

Created this classifier to help hospitals triage the suspects of COVID-19 from a simple blood test! It is faster and spares many SARS-CoV-2 test kits. It can save lives and costs simultaneously during the peak of the coronavirus pandemic.

What if we could speed up the diagnosis process? A novel classifier was trained with a dataset of X-ray images from children's chests. Now, it labels any x-ray image as either normal, bacterial pneumonia, or viral pneumonia. The result is an automated diagnosis that can accelerate the clinical care for children with world-class precision.

Applied computer vision and digital signal processing techniques on the radar for classification. The novel model finds meaningful contrasts between weapons and decoys of a person in movement. CNN can be used to analyze images and any other signals such as video, audio, radar, and sonar.

What if artificial intelligence could kickstart an original text, so you do not have to start a new one from scratch? I created this website that the user enters a few keywords, and it gives back some related suggestions of text with titles. I used Heroku, RESTful, and asynchronous tasks to deploy this model online.

Based on a collaborative filter, I created a recommendation engine for a cloud-service online retailer. Any online business must leverage its database to find out and deploy a better user experience to drive consumption. Unsupervised learning is ideal to find customer segments, which can be used to predict common preferences.

The world's busiest airways passengers now know if their flight will delay with 15 minutes of margin. I created this supervised learning model, considering a 2-year long hourly dataset of airport departures, arrivals, and weather.

The Parrot quadcopter just gained an autopilot. Thanks to thousands of simulated runs, this new reinforcement learning knows how much power each rotor must apply to stabilize the drone. Let me know if you need specific maneuvers or autopilot for any other vehicle.

Now, the owner of a bike-sharing company from Washington can make better decisions on how many bikes to buy. This forecast helps the owner to avoid losing revenue by having too few or skyrocketing the operational costs by having too many. I created this model by training an RNN over a 2-year hourly rental and weather data. RNNs are suitable to predict any time-series data such as demand, weather, and stock options.

A retailer from Portugal wanted to lower the frequency of deliveries to reduce costs without considerably disturbing its clients. It was achieved by finding customer segments and selecting which would not be significantly impacted by three weekly deliveries instead of five. Unsupervised learning is ideal for finding customer segments and understanding their specific purchase behaviors.

The style transfer technique allows the convolutional neural network to learn the content independently from one picture and the style from another. Later, I gradually applied the style from the painting to the content of the picture with the octopus.

Designed, embedded, and tested in the field with telemetry during real launches algorithms for missiles and smart-bombs, including aided navigation (GPS, BaroAltimeter, and IMU), transfer alignment, sensor fusion, target position estimation, and extended Kalman filter.

Created a predictive model to read a movie review and classify it as POSITIVE or NEGATIVE. The model applies text processing techniques and a neural network. This technique is also reusable for other applications such as product reviews or stock market discussions.

Online advertisement opens room for custom experience based on the user's known preferences. This opportunity is explored by offering an automated way to change the image on website banners with the brand based on which user is accessing the website.

This methodology is achievable by reusing a trained ResNet101 (transfer learning) for image segmentation, followed by image processing to find the best spot on the background to place the logo.

Structured the technologies for a new virtual showroom startup.
Planned the scope of the proof-of-concept (POC) and minimum viable product (MVP) for a startup about to launch a new mobile shopping experience by combining augmented reality, photogrammetry, artificial intelligence, simulations, and virtual try-on.

The outputs are a photorealistic image of the user wearing the virtual garment and a heatmap as a realistic representation of how the garment will fit and the proper size for the specific user.

Proposed new state-of-the-art solutions (better than photogrammetry) to scan the human body with great accuracy.

I developed the video processing technology core for one of the big five giant American tech companies. Now you can use your camera to create custom special effects that rely on background removal and augmented reality animations. This technique is transferable to online streaming data, such as filters.

A series of studies to find opportunities for Forex trading.

The studies included:
• analysis and feature engineering of historical real trading data.
• a custom-trained model with emphasis on explicability.
• a custom Kalman Filter for intraday.
• time-series ARIMA analyses.
• autocorrelation.
• pair trading.