Miguel Tasende

This was the capstone project for Udacity's Data Scientist Nanodegree, addressing Starbucks' challenge to select the best offers to show to each user in the cellphone app. The project is based on simulated data from Starbucks.

Starbucks has three different offers for their customers that use the mobile app:
• BOGO, buy one get one, which allows the customer to get a free product when making a purchase, with a specific duration
• Discount that enables customers to purchase the product at a discount for a given period
• Informational that shows ads to the customer

The project aimed to find the best offer for each customer to maximize offer completion probabilities or profits. Only one product would be considered per customer.

Medium story about the project: https://medium.com/@miguel.tasende/starbucks-offer-optimization-adb323ca32b5

An automatic dashboard to track the USD/UYU pair. The charts show the USD/UYU exchange rate and estimates of the expected value according to the purchasing power parity (PPP) theory. All estimates were calculated causally based on present data; no predictions were made. The data source is the World Bank.

GitHub repository: https://github.com/mtasende/usd-uyu-dashboard

A software to achieve the multi-target tracking of worms in a water drop for a biological research application. The goal was to detect worms in a video in which they were moving in a drop of water. The water contains elements that should be rejected by one kind of worms while ignored by others. To prove that hypothesis, it is helpful to count the number of worms in the droplet.

The project was meant to predict stocks' close and future date values, given a time series of open, high, low, close, and volume values for the stocks included in the S&P500 index in January 2017. Future date predictions referred to anytime after the dates used as training data. The project also considered the needed period of data for the prediction and the expected prediction accuracy in future dates. It was inspired by the assignments of Udacity's Machine Learning for Trading course and its corresponding course at Georgia Tech.

An automatic trading system was also implemented as a secondary problem, using information from the previously trained predictor at least in one of the versions. This system's goal was to maximize profit within a specific time horizon. Available values were considered to decide whether to buy or sell some equity and in which quantity, assuming it is possible to do so at an approximate close price.

Parallella is a hybrid computing platform built from the Kickstarter project held by Adapteva. It is composed of the high-performance, energy-efficient, manycore architecture, using an Epiphany chip as a co-processor and one Zynq-7000 series chip that usually runs a regular Linux operating system version. It serves as the central processor and implements glue logic in its internal field-programmable gate array (FPGA) to communicate with the many interfaces in the Parallella.

This project refers to an Epiphany-accelerated BLAS library created for the Parallella platform, which could also be suitable for similar hybrid platforms that include the Epiphany chip as a co-processor. Used the BLIS framework for the actual BLAS instantiation. Previous implementations of matrix multiplication on this platform have achieved outstanding performances of up to 85% at peak inside the Epiphany chip but not so good performances for the complete Parallella platform due to inter-chip data transfer bandwidth limitations. The main purpose of this work was to get closer to practical linear algebra applications for the entire Parallella platform with scientific computing in view.

Github: https://github.com/mtasende/BLAS_for_Parallella