Transformer Models
Multimodal Toxicity Modeling in Commentary Media
I built a multimodal deep learning system to measure how audio signals influence perceived toxicity in political and social commentary. Using short video segments from Ben Shapiro, Joe Rogan, and Jon Stewart, I constructed a pipeline that combines transcript text with corresponding audio clips to predict a continuous toxicity intensity score between 0 and 1.
The project compares a text-only baseline built on RoBERTa with an audio-augmented model leveraging Wav2Vec2 embeddings, isolating how vocal features such as tone and prosody shift toxicity judgments relative to text alone. To scale labeling, I designed an LLM-assisted weak supervision framework that generates noisy toxicity annotations, later refined into a high-quality evaluation set. Results show that incorporating audio meaningfully alters toxicity predictions, highlighting the importance of multimodal context when modeling harmful or aggressive language in real-world media.
Populism Detection & Summarization
I developed a transformer-based model that analyzes political press releases to both summarize content and detect populist rhetoric. Using GPT-4o on a dataset of 10,000 releases from 12 European countries, I engineered prompts to label where populist framing occurs—such as “people vs. elites” or anti-institutional language—and to generate concise explanations of why certain phrases qualify as populist. The final model, built on a fine-tuned BART architecture, allows users to quickly identify when populism is present in political discourse and understand the reasoning behind that classification.
Text Summarization
I built a transformer-based summarization model using T5 to generate concise, abstractive summaries of political press releases. Trained on 10,000 releases from 66 parties across 12 countries, the model leverages prompt engineering to capture the core message of each document while filtering out less relevant details. This enables users to quickly understand the key points of complex political texts and compare messaging across parties and nations.
Research Papers
(Code Available Upon Request)
Randomized Search Techniques for Real-World Optimization Problems
This project examines two everyday scenarios through the lens of optimization by applying three local randomized search algorithms—Randomized Hill Climbing (RHC), Simulated Annealing (SA), and Genetic Algorithms (GA)—to two distinct optimization problem domains that mirror real-life events.
Next, I extend our exploration by using these algorithms to optimize the weights of a neural network, building on a problem space from previous work.
Exploring Political Contributions Through the Lens of Machine Learning Algorithms
This paper analyzes data from the Federal Election Commission (FEC) and the Wall Street Journal’s comprehensive school shooting database to develop four predictive models using the following machine learning algorithms:
The first dataset employs Support Vector Machines (SVM) and Extreme Gradient Boosting (XGBoost) to predict political contributions to Democrat or Republican-affiliated committees, based on a range of socioeconomic and donation-specific variables. The second dataset applies Neural Networks and K-Nearest Neighbors (KNN) to examine the influence of school shootings on donations to pro-gun and anti-gun committees.
Exploring Reinforcement Learning Algorithms in Deterministic and Continuous State Spaces: Blackjack and Cartpole
This paper analyzes interesting properties involved in the Markov Decision Process (MDP) through the analysis of two classic reinforcement learning environments as part of the gymnasium library: Blackjack and Cartpole. By comparing model-based approaches such as Policy Iteration (PI) and Value Iteration (VI) and model-free approaches such as Q-Learning, this work investigates the computational efficiency, convergence behavior, and policy outcomes across deterministic and continuous state-space environments.
Machine Learning For Trading
Developed a comprehensive AI-based stock trading system. Built key components of a simulated high-frequency trading architecture, including order processing, trade execution, portfolio accounting, market simulation, and strategy evaluation. Integrated supervised learning models (XGBoost-based Strategy Learner), reinforcement learning agents (Q-Learner), and custom trading indicators to optimize trading performance. Designed and tested the system iteratively, culminating in an agent that executes a trading stragegy using advanced machine learning techniques with realistic market data and trading constraints.
Re-Imagining Data Collection and Analysis in Special Education:
A research paper culminating in the development of a medium fidelity prototype focusing on the redesign of PowerSchool’s Individualized Education Programs (IEPs) interface, particularly the Present Level of Performance (PLOP) interface.
Yummly Advanced Ingredient Search:
A research paper focused on refining the ingredient search tool for the Yummly Web Application through several iterations of the design lifecycle.
Desktop Applications:
Green Collectives - Desktop Application
Because of my deep ties to the military, I’ve always had a strong desire to support fellow servicemen and women. During my time as a Green Beret, I had the privilege of assuming leadership roles in some of the world’s most politically sensitive and complex regions, spanning from the Middle East to Eastern Europe. This is what drove me to create an all-in-one solution—an application designed to furnish soldiers with convenient access to essential files, a comprehensive directory of military websites, resources for constructing and overseeing shooting ranges, engaging workout routines, and real-time updates on current events.
If you prefer a high level overview of the project, check out the README File.
You can check out the application from the Snap store here:
Web Applications:
Green Collectives - Web Application:
Another project I completed addresses a persistent challenge faced by soldiers — limited access to essential files during long training cycles, permanent changes of station (PCS), or critical system updates. To solve this, I built a secure, collaborative file-sharing platform where service members can upload and access unclassified resources anytime. The site is structured around an HTTP API Gateway that routes requests to a Node.js Lambda function in AWS, which generates pre-signed URLs for secure GET and PUT operations to an S3 bucket. Click here to see the site.
Mobile Applications:
Android Job Comparison Application:
A job comparison application built in Android Studios using Android Room.
Terminal Programs:
Propositional Logic / Logic Circuit Calculator:
A C++ program that takes a series of logical connectives as user input and displays a truth table as binary output.
Logical connectives are pivotal operators that amalgamate or modify logical statements to construct intricate propositions, enabling the articulation of relationships among these statements. These are indispensable in building logical expressions and in analytical reasoning. Truth tables are essential tools in visualizing and resolving logical dilemmas, engaging in Boolean algebra, assessing logical equivalence, and facilitating decision-making processes. Boolean operations underpin many everyday functionalities, including search engine queries.
Users input a sequence of logical connectives, and the program renders these inputs either as truth tables or as logic circuit diagrams. A suite of algorithms will parse and iterate over characters, variables, and constants to produce the desired outputs.