Things I've created.

This started as my final project for an LLM applications course at school and quickly became the most technically involved thing I've built. SpielBot uses hybrid dense/sparse retrieval with reciprocal rank fusion combined with multi-query reasoning and a vision pipeline for interpreting game-state images. I benchmarked it against commercial competitors like Board Game Wizard and RulesBot.ai using a custom 30-question evaluation set scored by an LLM judge. More to come!

Our take on the company brain problem, tackling the fact that companies lose millions a year when experienced employees retire or leave and take decades of tacit knowledge with them. Our product is a voice-first work journal that just asks people what they worked on, then rewards them for answering, turning that knowledge into something the whole org can draw on to support and grow its people.

Working with Massa Baali and a team of student researchers, we picked apart CoLMbo-DF, an audio language model that reasons over acoustic features (pitch, formants, HNR) to detect deepfake speech. We ran statistical analyses across ~15,000 speech samples to qualify signals and noise in both authentic and synthetic samples. One fascinating find was that the model's own reasoning would correctly call out synthetic artifacts and then still land on an incorrect classification, highlighting a disconnect about where modern speech systems break down.

Physicians in concierge medicine settings are buried in EHR data and need to surface what actually matters for each patient visit, while patients have no meaningful way to interpret their own medical records. Our solution was a dual-sided platform with both a physician-facing dashboard that uses AI to generate pre-visit summaries, flag trends in labs and vitals, and give doctors a "board-level" view of each patient, and a mobile app that translates clinical notes and lab results into personalized care plans the patient can actually interpret.

A more involved data analysis project diving into the international wave of mass layoffs at the time, with a particular focus on the tech industry. Our team was curious to understand what was actually driving them from a statistical lens by pulling together layoff data across industries and time to uncover significant patterns in the economic story behind the job market.

My friend William and I had two goals here: understand what drives output quality for neural style transfer and apply those findings to improve transfer across various style expressions. We picked out Gatys et al.'s classic VGG-based model, running over a hundred trials across hyperparameters. After testing candidate combinations of those findings, we managed to design a revised algorithm that beat the baseline on both loss metrics and visual appeal. This was a tricky one to balance since lower loss didn't always mean a better-looking image, but beauty is in the eye of the beholder I suppose!

One of my bigger projects with R, I performed multiple analyses of biomarker data from 191 subjects to identify which cannabis compounds and cut-offs best signal recent marijuana use. The study compared candidate matrices and thresholds with the goal of evaluating arbitrary defaults used in U.S. legistation against reliable clinical data.

As an attempt to practice my ML programming, I implemented core supervised learning models from scratch in Python and benchmarked them across several datasets. From the experimentation, I discovered many aspects of where each model shined, broke down, and how these implementation details influence the model selection process.

My team and I analyzed a decade of American commerce data against pandemic infection rates to see how 18 major U.S. industries responded to COVID, then built regression models on the standout sectors to quantify the impact and separate the industries that seemed to have bounced back fast from the ones that didn't. Admittedly, this analysis was conducted too soon to meaningfully capture the complete impact of COVID-19. I'd be curious to run it once more...

We built and compared two predictive models on CDC survey data from 250,000+ people to determine the strongest predictors of heart disease. Testing a multiple logistic regression model against a principal component approach, the former came out on top at 77% accuracy. While we'd guessed smoking and diabetes would be big drivers, general health rating and patient sex also ended up landing in the top five features.

This was my first stab at a computer vision/deep learning project! My partner and I put together a CNN inspired by BrainMRNet for classifying fMRI scans into four brain-tumor categories. The project covers preprocessing choices, architecture decisions, and training dynamics that pushed test accuracy to 93%.