Iain Ramirez

Designed and executed a large-scale study evaluating the reliability of large language models across 60000+ synthetic documents, systematically analyzing how factors like noise, length, and information placement impact extraction accuracy in particular for the use of automatic database construction. We are aiming for a Co-authored paper submission in June of 2026 to the Journal of statistical software. The project is part of ongoing research I perform with Hudson Golino, whose work focuses on combining machine learning and quantitative psychology to develop new methods for analyzing complex data. My research was recently accepted for presentation at the International Meeting of the Psychometric Society (IMPS 2026) in Seoul, South Korea, one of the leading conferences in quantitative psychology and psychometrics. Additionally, the team as a whole was featured by UVA in a published article highlighting our team’s approach to “reading artificial minds,” showcasing how human insight can be used to evaluate machine intelligence. This project is still underway, when complete a paper will be attached. Photo credit: Evan Kutsko

I conducted experimentation on single-cell B-cell datasets using the ProtoCell4P framework, running 4,000+ model configurations to evaluate performance across hyperparameter settings. This work focused on improving model accuracy and robustness in high-dimensional biological data. I used heatmaps to identify high-performing regions of the hyperparameter space. These insights directly informed improved configurations and training strategies, improving ROC-AUC from 0.51–0.70. To enable experimentation at this scale, I used the UVA High Performance Computing (HPC) cluster and developed Slurm parallelization scripts, significantly accelerating model training by up to 97% and allowing for efficient hyperparameter sweeps. Worked closely with Professor Stefan Bekiranov at the University of Virginia, contributing to ongoing research at the intersection of machine learning and computational biology.

I designed and deployed full-stack websites and automation systems for 7+ organizations, including student groups and small businesses, supporting 3,000+ users and generating 10,000+ site visits within the first month across projects. I built custom platforms for scheduling, registration, and content management that replaced manual workflows and saved teams 4–6+ hours per week. I developed these using PHP, JavaScript, SQL, and TailwindCSS, focusing on making them simple, fast, and easy to use. I worked directly with clients from idea to deployment, turning their needs into working systems and improving them over time based on real user feedback. Some of the organizations I’ve built for include UVA Club Grappling, UVA Muay Thai, UVA Girls Who Lift, and UVA Plushies for Patients. Additional project links (not listed here as systems have since changed) include Mountain Mama Yoga, The Golf Mission Tour, and SFS Baseball.

FighterForecast is an machine learning system for predicting UFC fight outcomes using historical data and fighter performance analytics. This project is built on a dataset of 8,000+ fights, where the model incorporates 40+ features capturing fighter momentum, recovery time, and skill-based attributes. The system uses a stacked classifier with multiple models (logistic regression, random forests, gradient boosting) with a focus on out-of-sample generalization. Through feature engineering and modeling improvements, the project achieved 66% prediction accuracy (vs. 50% baseline). To evaluate real-world applicability, I integrated a Kelly criterion–based betting strategy, simulating 10%+ annual ROI, and deployed the model as a full-stack web application with automated data pipelines and live predictions. This project got 20000+ views on social media and 200+ followers.