About Me

"Mathematics is the most beautiful and most powerful creation of the human spirit," Stefan Banach once beautifully stated. As a young boy, I was fascinated as I learned about mathematics and its application in the world. It soon became something that I yearned for as it helped me analyze solve problems with precision, rigor, and critical thinking.

I am Oscar J. Escobar and am an applied mathematician. I love football (i.e. soccer) and am a huge fan of Real Madrid. In my spare time, I like to cook, watch series and movies, play the piano and guitar, and learn new things.

I have found a true calling in taking complex real problems and turning them into bite-size pieces of math. I then bridge the gaps of what is yet unknown with a precise solution and quantitative analysis. This has often helped me earn the praise of being an innovator by using math to model the world and find solutions.

My passion for solving real world problems has led me across several industries and research projects, one of which I was able to publish. My research has been in the field of engineering in the computational fluid dynamics as well as wildfire model optimization. I also work with reinforcement learning in modeling attacks on electrical infrastructure. For industry, I have leveraged mathematics to employ statistical analysis and machine learning to asses or build models in the Natural Language Processing area as well as image text extraction.

You can checkout my current and past work/project results here.

Skills

Here is an overview of my skills. Feel free to also look at my résumé.
  • Optimal Control Theory: PMP, HJB, LQR & LQG
  • Data Assimilation: Kalman, Particle Filters, EnKF
  • NLP
  • Deep Learning (PyTorch & Keras)
  • Deep Reinforcement Learning: DQN & Policy Gradients
  • Time Series Analysis & HMMs (statsmodels & hmmlearn)
  • Machine Learning (SciKit Learn)
  • Bayesian Statistics & Probability Theory & MCMC
  • Python & SQL

Education

I am currently pursuing a Bachelor's of Science in Applied and Computational Mathematics (ACME) from Brigham Young University. My anticipated graduation date is in Dec 2025. Some of the important and relevant coursework I have taken so far is:

  • Mathematics of Deep Learning
  • Data Assimilation
  • NLP
  • Modeling with Data & Uncertainty
  • Optimal Control
  • Deep Learning
  • Optimization
  • Modeling with Dynamics (ODE & PDE)

Recent Professional Experience

You can find a full list of my other internships on LinkedIn. You can checkout my repo where I host my projects.

ML Intern Developer | June 2024-Aug 2024

I was able to work at Wells Fargo under the Analytics & Data Undergraduate Program (ADUP) in Minneapolis. ADUP gave me the opportunity to choose the sorts of projects I wanted to work in, and having had experience in ML from my previous role, I chose ML modeling and data analysis. Given the choices, I chose to go into the Advanced Data Analytics & Solutions team under Andy Chandler to work more on the NLP opportunities his team presented.

The Advanced Data Analytics team was looking at data from previous Wells Fargo projects. This data contained information about timelines, next steps, people on the team, description of the project and scope of it, as well as other information like dollar amounts and people who get covered by the project. But, the data was very messy and inconsistent as measurements, collection, and organization is concerned. Our main task was to answer 2 questions: 1) Does the given dataset possess anything that is predictive for assessing project risks and timelines? 2) Given the assumption of 1, can ML or NLP help assess project risks and timelines?

As per the results, in answer to question 1, I was able to find predictive power within the text description of the business projects, particularly in the risk descriptions of the project. More often than not, the risk description was pretty predictive in helping give a binary classification if a given project can meet Wells Fargo's established timelines (an anwer to 2). I was able to give the team a trained NLP text classification model that employs an SVC classifier. Due to the volatility in the data, the classifier had a performance range of 63-72% classification accuracy for binary timeline prediction. But, the model did better, with a performance range of 73-79% classification, in classifying projects into the different Wells Fargo risk tiers. My code also took care of the text cleaning and standardization where I eliminated stop words and use 2-grams.

ML & Data Science Student Engineer | April 2023-April 2024

I was able to work in the Family History Center, located in Lehi, UT, of The Church of Jesus Christ of Latter-day Saints. During my time at the center, I was able to work with a machine learning production model that extracts genealogical information from digital artifacts. The main goal was to improve the model's ability to correctly identify crucial information from the digital artifact that was needed to help create more online records in less serviced populations like Spanish speakers, Portuguese, and Hungarian.

I initially helped identified error and improvement areas (primarily being name identification and spelling) of the model that helped increase the performance by around 20-30%. To do this, I created my own Python script that would take the genealogical information output, as a .txt, and compared it to manually extracted and curated information (i.e. data labels). The script would then do various comparisons, like searching for biggest contiguous string, and perform basic statistical analysis to help identify areas of improvement.

Furthermore, I also led efforts and meetings with a team of 3 data labelers in creating the data labels to help train the model. My other responsibility was also documenting our findings and procedures to help further define the process of data labeling and data analysis of our team and role. Because of how well I did, my manager, Jon, offered me a part-time extension towards the end of my internship that I could continue as I went into my then junior year of the ACME program. I happily continued the role and fulfilled the same responsibilities until the end of the internship.

My Work

You can checkout my repo where I host my projects. I give a more detailed explanation of my current research and projects here. In my repo, I host more of my code and results.

Research

Wildfire Model Optimization; BYU - Missoula Fire Lab (Sept 2024-Present)

I also started this other project with Dr. Barker from the mathematics department in conjunction to his work with Missoula Fire Lab. Wind is a big principal component of wildfires, but it is hard to model properly. WindNinja is a software that uses two different models that can solve the wildfire ODEs in order to make a wind profile. However, one method is really fast but inaccurate, whereas the other is accurate but very slow. The main goal of the project is to find a way to perform spectral methods on current wildfire modling techniques where we can implement the Fourier Basis. The next goal would be to implement the Fast Fourier Transform so that computational cost is truncated significantly with the hope of being able to run some of the more complex models in laptops of forest agents. This would allow us to make wind modeling fast and accessible at the time of disasters. I presented my current work here .

Optimal Control in Mathematical Oncology; (Sept 2024 - April 2024)

As part of my Modeling with Dynamics class, we have to perform a small research project where we get to select a specific system or world phenomena we would like to explore and model. My team chose to model Breast Cancer growth under the effects immunological response and chemotherapy. We first focused on finding an ODE model that incorporated tumor-immune-cell interactions modeled by Michaelis Menten Kinetics (MMK). I did the main research for both models and came across Gompertzian growth as well as an exponential decay of the chemo (mainly by reading DePillis).

The next phase was to pick a specific cancer, invasive ductal carcinoma (IDC), and modeled that with a specific treatment. We opted for the Adriamycin-Cytoxan (AC) treatment which comprises doxorubicin and cyclophosphamide. I did the main application of optimal control theory in order to find an optimal continuous dosage. I leveraged Pontryagin's Maximum Principle and used an iterative method to numerically solve for the continuous dosage. We were able to find an optimal dose that seems aligned with existing research of bang-bang solutions. However, our in-silico results show a continual decay of epithelial cells which made us question the accuracy of our model (at least in the chemo effects portion). If you would like to see our paper, please contact me.

(Published) Nozzle Design using CFD; BYU Rocketry Club-Hybrid Rocket Research Team (Sep 2021-Aug 2023)

I joined the BYU Rocketry Club: Hybrid Rocket Research Team back in the fall of 2021. We were the recipients of a grant from the Utah-NASA Consortium of about $8,000 that was to be used to generate research for paraffin LOX hybrid fuel. We created a design for a 12ft rocket that would reach a target of 10,000 ft and have a maiden voyage in the summer of 2022. We (see below) were able to present our findings at the Utah-NASA Consortium 2022 Spring Conference and publish it in the Utah-NASA Consortium webpage .

I was working within the propulsion subteam to create the LOX dome, combustion chamber, fuel grain, and nozzle. I was tasked with implementing a design for the nozzle configurations that would enable us to reach a target exit mach number of 3.1. I studied the works of Gas Dynamics by Zucrow and learned more about the method characteristics employed for nozzle design. I was able to implement a Python3 script that would similate axisymmetric irrotational compressible flow given a throat area, radius of curvarture, desired exit mach number, and specific heat of the fluid. The script would output two solutions: 1. Rao nozzle in 3D (using Matplotlib) and 2. a txt, csv, excel file of the nozzle characteristic points that would then be used to create a CAD model in SolidWorks. I was later able to implement numerical methods for the transonic region analysis. The main goal of this was to better estimate the mach number achieved by the nozzle as well as better configure the mesh of the method of characteristics.