Manuel Montoya Catala

A Dash-based app for factory monitoring and optimization. I served as the senior data scientist leading a team of four people where we implemented ETL pipelines, data analytics, and dashboards. They were used for recipe and linage optimization, as well as for monitoring the factory equipment.

Trapyng is a framework designed to build automated trading systems in Python, allowing easy development, testing, and real-world usage of trading ideas. It intends to bring proper mathematical understanding to current trading strategies used by retail traders and to apply data science in order to improve them.

The platform focuses on the proper visualization of the strategies and the understanding of the underlying methods, promoting the generation of new ideas for future trading strategies.

I started this project in 2016 with the goal of properly applying machine learning in trading. I took professional trading courses and opened a trading account with a broker to learn how retail traders think and use those insights in the system and data modeling phase.

A few manuals have been created describing the features of the system, although a large number of them are still undocumented due to lack of time. These include advanced linear models, Kalman filters, Gaussian processes, and RNNs, among others.

B-ADSN is a neural network architecture that aims to combine the expressiveness of Deep Learning, with the robustness to over-fitting of Boosting.

This project describes the B-ADSN architecture developed during my research at UC3M. The main results of this research can be found in the referenced IEEE Xplore Publication and all other materials including codes can be found in the linked repository.

The code was entirely developed from scratch in Python without TensorFlow or any other automatic differentiation library since they were not as popular at the time of development in 2015.

A tool that allows the modeling of data using multi-modal time-dependent directional distributions. It can handle combinations of the Watson, Von Misses Fisher, and Gaussian distributions. It employs the expectation maximization algorithm both for independent clusters and for Hidden Markov Models of order 1.

The tool was created from scratch in Python and It contains many functionalities such as being able to combine clusters from different distributions, it can recover from degenerate clusters and it is very customizable.

The code of the final tool can be found in the referenced repository and along with its documentation. In the repository, you can also find example videos that can be generated by the tool and its applications to EEG data.

In this project, soft-computing techniques are used in the prediction of the sea-wave characteristic of specific areas of the Caribbean Sea. The main goal is to identify the principal predictor parameters of the wave-height of a given location.

Neural networks are used in order to predict such height, the input data is a combination of historical weather conditions and sea-waves parameters from other areas of the ocean worldwide. Meta-heuristic search techniques are used to find the subset of these parameters that produces the best prediction.

The project is completely implemented in C from scratch using the CBLAS library and POSIX threads for optimizing the speed.

The main focus of the project is the feature selection of the variables, which is done by combining a fast regression algorithm, the extreme learning machine, and discrete optimization local search heuristics such as simulated annealing and evolutionary algorithms.

This Python class provides the core structure of a Bayesian BiDAF-ELMo system implemented in PyTorch for the question-answering (Q&A) task.

The architecture incorporates multiple deep learning models, including bi-directional LSTMs, highways, and attention mechanisms. Additionally, I have extended these architectures with a Bayesian approach, which can be accessed in the repository.