Projects

Django, SentenceTransformers, UMAP-Learn, Python, JS, HTML/CSS

Developed a full-stack Django web application into production hosted on Sevalla with a PostgreSQL database. App clusters and visualizes card text using SentenceTransformers and UMAP dimensionality reduction.

Love2d, Lua

An interactive sandbox toy with time reversal mechanics.

FastAPI, Python, HTML, Javascript

Developed a full-stack data loss prevention framework through a Chrome extension with a FastAPI backend. LLM-based agents increased redaction scope and accuracy over traditional regex methods. Communicated with company leadership to scope and deliver the finished product.

Pytorch

Compared CNN, RNN, and LSTM hybrid models for keystroke decoding from sEMG using Stable-Baselines3. Found that the baseline CNN model outperformed the hybrid models, achieving a character error rate of 21.82.

Pytorch, Stable-Baselines3, Metadrive

Used proximal policy optimization to train autonomous driving agents in Metadrive environments. Maximized route completion while tuning hyperparameters like scenario count, clip range, and reward shaping. The best agent achieved 88% route completion and 70% success rate.

Pytorch, CREST

Investigated how coreset data selection effects lottery ticket one-shot neural network pruning. Trained ResNet and LeNet models on CIFAR-10 and CIFAR-100 datasets over hyperparameters including post-pruning epochs and prune-percent. The fine-tuned models maintained 2% higher accuracy after 10 epochs, suggesting coreset selection reveals structural patterns.

Python, Streamlit

Built a predictive web app using Streamlit and Scikit-learn to estimate IMDb ratings, enabling users to input show attributes and receive a predicted popularity score. Scraped and parsed thousands of entries from Kaggle and IMDb, and engineered features to train a random forest, KNN, decision tree, and beta regression with an RMSE ≈ .197.

R

Applied autoregressive models such as SARIMA and TAR to global temperature data spanning 100 years. Forecasted global mean temperature over the next 80 years to identify climate trends.

Python, SQL

Used hyperparameter tuning and cross-validation to compare an Elastic Net, Decision Tree, Random Forest, and Boosted Tree in predicting Super Smash Bros. match outcomes. Highlighted that stage selection and player performance history were more predictive than character choice alone through variable importance, achieving an accuracy of 67.8% with a Random Forest model.

HTML, CSS

A "Wordle"-like web game with a twist.