Miguel Cozar

EXOS

An agile end-to-end consumer digital platform for lenders

EXOS Technology is a digital solution offered by ServiceLink. It provides an agile end-to-end consumer digital platform that lenders can utilize at any point in the digital process, enhancing every critical consumer touchpoint from origination all the way through servicing1. Here are some key aspects of EXOS Technology:

Origination Services: EXOS Technology streamlines and eliminates inefficiencies in the mortgage process. It offers services such as payment processing, consulting, and software development.
Borrower-Centric Technology: With EXOS, borrowers experience speed and transparency throughout the mortgage journey. They can instantly schedule appraisal or closing appointments, enhancing their overall experience.
Cloud-Based Digital Solutions: EXOS provides cloud-based digital technologies to real estate lenders and servicers. It transforms consumers’ digital expectations around the mortgage experience through mobile apps, voice interaction, APIs, Predictive Analytics, and AI

My main work is on the AI side, where we have developed several Azure based models capable of extracting information from unstructured documents on a structured way. I have used Sequence to Sequence models, LLMs, Computer Vision models and multi modal LLMs. During my stay at ServiceLink, I have also contributed with optimization algorithms such as Genetic Algorithm. I have also impulsed new work procedures, such as an automatic agreement score calculator that has boosted our quality labeling for model fine tuning. Visit Exos Website

Forex Platform

An AWS hosted automatically deployed platform for currency data

The objective of the whole project is to develop a cloud hosted trading platform. I developed a Stream Data Collector, implemented a time series database and a visualization server. All of this services run in Docker containers in an EC2 instance automatically deployed with Terraform scripting. Links of the project:

The project is designed to have a bigger platform, including more instances. Some tests are also provided. The platform supports more than 50 currencies and returns stream graphs of the price evolution.

Tests

Platform Global Architecture

Developed Architecture

Song Popularity Prediction

A Deep Learning project with Convolutional Neural Networks.

The main problem is to estimate how popular a song is going to be according to its features. These features to be included are distributed in three main groups: general information about the song, music features analysis, and lyrics analysis. We designed three convolutional neural network architectures to try to detect what's important for a song's popularity. Based on our data, we detect that lyrics are quite relevant for the sake of this metric, and we extract a 3D figure of the lyrics, with PCA dimensionality reduction that helps to see that relevance. Our model gets to have a MSE of 0.0031 in lyrics testing, demonstrating that it is possible to try to predict a song's popularity. Resources:

These are some of the most relevant figures.

A 3d figure representing the songs lyrics colored by popularity

Lyrics Model Architecture

Model error distribution

News Summarization

A Twitter stream bot equipped with fine-tuned BERT summarizers to answer users' requests

Our project aims to develop a generative model able to autonomously generate concise and informative summaries for news in social media. After months of research and development, we present successful results that fulfil our initial objectives. Our generative model is capable of creating readable summaries that capture the essential points of news articles. To showcase the practical application of our model, we implement a real-world use case by deploying a Twitter bot that generates news summaries when provided with a URL. The bot processes the input text by employing the Flesch Readability Test and a "Legal" Bag of Words to filter out noise and generate legible summaries, answering the users' queries by citing the bot Twitter's account.

These are some of the most relevant figures.

Model Employed

Twitter Bot Architecture

A US Senate Data Analysis

A Data Analysis of the US Senate social media relations, with a publicly available big dataset

Social media, specifically Twitter, plays a crucial role in political discourse and communication. It is the window of voters to their candidates, and what they publish may determine their success in the elections. A deep analysis is needed in order to better understand the current situation and think of strategies to reach the audience. This project joins the creation of a self-made dataset, machine learning topic models, geography and a brand-new created popularity metric to explain the current political landscape and provide insights about the most influential senators and their discourse.
We also public all our results in a Kaggle public dataset available on the links below, alongside with a notebook that details our results and findings.

These are some of the most relevant figures.

Map of the most Popular US senators and their Twitter connections

Senators by Popularity and number of followers

Concierge

A hotel management application developed with React and Java, following agile methodologies.

Check our Hotel Management application. The front end was developed using React and we leveraged Java SpringBoot for the Back End. We created a database with SQLite and worked as a team following Scrum methodology, where I acted as Product Owner. Resources:

ELC

A regional Telecommunications provider, offering broadband, satellite communications, VoD...

We worked in teams of 16 members to develop a regional telecom provider. The services included broadband access with fiber optic, 5G mobile network access, IoT systems for agriculture and port management, VoD and satellite broadband access for isolated areas in the region. We worked on the design and the dimensioning of the services, the technical specifications and economic plan. Resources:

Project Presentation

Tweet Classification

A semi-supervised Tweet Classifier using density clustering and Bi-LSTM Neural Networks.

Final Bachelor project as a researcher on the Speech Technology Group at UPM. It was also my first Data Science project. I built a semi supervised tweet classifier using HDBSCAN, UMAP and TF-IDF to cluster the tweet dataset. I later used the clusters to get an approximation of the topics, based on key words. Finally, I leveraged a Bi-LSTM Neural Network to classify the tweets using the weak labels. I also include some visualization and statistic modules. Resources:

Project Paper

Statify

An analysis of the cultural differences on Music Tastes.

Have you ever asked yourself what needs an artist to be on the Global Top? Is it the same for every country? How does countries culture influence artists success? In this project, I try to explain all these questions using a Python script (publicly available) and data provided by Spotify API.
I use the song features provided by Spotify: danceability, acoustics, key, speechiness, liveness, energy, valence, and instrumentals. In addition, I created a new interesting metric called versatility, a combination of the prior with unsupervised learning. Basically, all the features in an array form a point in an eight-dimensional space (a matrix of 8 items). If we use all the tracks of an artist, playlist, or whatever, we obtain a map of points that can be clustered and represented in 2 or 3 dimensions (in this case I have used HDBSCAN and UMAP). If the result is a small number of clusters, this indicates that the singer does not have a variety of registers, he would be monotone. Whereas in the other case, he/she would be considered to be versatile, varying their tracks frequently.

These are some of the most relevant figures.

Kanye West Versatility Metric

Tyga Versatility Metric

Tests

Platform Global Architecture

Developed Architecture

A 3d figure representing the songs lyrics colored by popularity

Lyrics Model Architecture

Model error distribution

Model Employed

Twitter Bot Architecture

Map of the most Popular US senators and their Twitter connections

Senators by Popularity and number of followers

Kanye West Versatility Metric

Tyga Versatility Metric

Published papers:

Registered Software Tools:

Programming Languages & Scripting

Machine and Deep Learning

Cloud Computing

Databases

Data Science Essentials

Other Skills

Contact Me: