Projects
All my machine learning work in one place.
Padel Match Analytics Pipeline
End-to-end CV pipeline turning one fixed broadcast camera into per-player padel match analytics.
Popularity Bias Debiasing in RecSys
BPR recommender on MovieLens 25M that cuts popularity bias 30% for under 10% accuracy loss.
LLM-as-a-Judge Reliability Auditor
Bias audit of 5 LLM judges, plus a calibration that lifts tie-detection 3–4×.
Rich, Fair, or Rich and Fair: Gender Bias Mitigation in Image Captioning
Inference-time demographic-fusion module that cuts gender bias in image captioners while retaining 97–99% of caption quality, with no retraining.
SayQL: Natural Language to SQL with GPT-Neo
Fine-tuned GPT-Neo 1.3B on Spider to turn plain English questions into SQL, lifting valid-query rate from 10% to 90%.
For fun
Side projects for the people I love, created to nurture our relationships and help them grow.
누리 (NuRi): Digital Confidence for Korean Elders
A mobile app for Korean elders in America. They were brave enough to start over here once, but the digital world has changed so fast they're being asked to do it all over again. It walks them through the everyday things in Korean, one short lesson at a time: getting around their phone, spotting scams and AI fakes, picking up more English, looking after their health, and a few brain games just for fun.
Friends Forever: Automated Friend-Group Newsletter
An automated monthly newsletter that keeps my friend group connected. Everyone fills out a Google Form; on the 1st, an Apps Script pipeline turns the responses into a magazine-style PDF and emails it to everyone. Custom layout and photo grids, with the Claude API summarizing some of the form answers along the way. Google Apps Script, Sheets and Forms, HTML/CSS-to-PDF.