Martin Baigorria

Currently working in the Adaptive User Interfaces team, working on a contextual bandits algorithm to optimize the layout of our mobile application. Refactored, automated and improved a recommendation engine for financial products on web and mobile platforms, achieving a 13% increase in BTR. This engine is also currently being used to optimize outbound communications with clients in order to maximize revenue.

Built a proof of concept for the product identification problem using image and title embeddings. Programmed our RESTful API to serve models and approximate nearest neighbor (ANN) searches on our product and metadata embeddings. Developed a ranking layer by predicting the conditional purchase probability of an item using gradient boosting (LightGBM). This achieved a 20% increase in buy through rate (BTR) in our view to purchases (v2p) recommendations over our previous solution.

Worked on the development of our recommender systems based on product embeddings (meta-prod2vec) for our views to purchases (v2p), purchases to purchases (p2p) and frequently bought together (FBT) items algorithms on the e-commerce platform. Achieved a general 35% increase in F-score, which translated into a 15% increase in BTR in v2p and p2p. Achieved a 2.5x increase in FBT items ATC rate.

Implemented a Commodity Risk Management tool which has now been incorporated by the Center of Financial Research at UTDT. Developed an algorithm to extract the risk-neutral distribution of soybean and corn futures prices from live CME option price data to evaluate the price volatility risks in the market.

Econometrics