Matriculation Prediction

This project, built with a team of graduate students for our NLP capstone, aimed to predict college matriculation from student application essays. I led the effort, designing and implementing a deep hierarchical transformer model using DistilBERT with chunked input to process long-form text.

The output probabilities were ensembled with a traditional Random Forest model and then aggregated using a meta-level neural network ensemble approach (MLP), producing a final classification model with precision, recall, and F1 scores exceeding 0.94. Additional confidence-based features from each model helped the ensemble reach an accuracy of 98.9% on the final holdout set.

I also implemented stratified cross-validation, class weighting, and extensive evaluation metrics.

Here is a link to our GitHub repository for the project.

Of the courses that I took during my Bachelor's degree. One of the most difficult classes was my Linux Programming course. In this class I did everything from recreating command line utilities from scratch using shell scripting, to socket communication, parallel processing, Operating Systems design, and creating my own version of the linux shell itself.

Here is a link to my GitHub repository containing all of my coursework.

Another partner and I developed a system for hiding copyright information within artwork images. We used Stenographic techniques to insert the copyright information into the image. This was done via making small encoded alterations to the RGB color values for pixels in one corner of the image. The result is a visually unaltered and unnoticable copyright tag that can be traced back to the original creator!

Here is a link to our GitHub repository for the project.

I was personally nominated to represent Sogeti in the Google GenAI Hackathon. Our team was tasked with creating an MVP (Minimum Viable Product) of a business application utilizing GCP's GenAI services. We used a webapp frontend plugged into a Python based backend hosted on Google App Engine, and communicated with Google's VertexAI chatbot model via their PaLM2 API. We created a personalized FoodQuest app which would find food options for users using their current location. The FoodQuest app was able to adjust for health conditions or allergies, or even different dietary restrictions. For example you could ask it for food options near you that are both vegan friendly and under 400 calories and it would provide a list of dishes at different restaurants that met the requirement.

Here is a link to our demonstrational video for the project.