Projects

A cognitive-governance kernel that makes AI agents show their work before they act. Before any high-impact action (git push, deploy, migration), a deterministic hook requires the agent to commit its reasoning to disk — knowns, unknowns, assumptions, and a falsifiable disconfirmation — and refuses to proceed until the artifact is real. Verified lessons become hash-chained, context-scoped protocols that resurface at the next matching decision, so the agent gets sharper on your codebase over time. Ships as a Claude Code plugin and a Python kernel with a vendor-neutral adapter layer.

A benchmark probing how well vision-language models understand human gaze, across four tasks: gaze target recognition, relative orientation reasoning, cross-view visibility estimation, and viewpoint-based accessibility. Built at MONET Lab (UIUC) from the MVGT dataset, with an evaluation harness that combines standard text metrics with an LLM-as-a-judge scorer so semantic correctness counts for more than exact match.

An event-management web application built and deployed for my battalion, processing 800+ participants across 32 events. It cut manual reconciliation by 75% and data-entry errors by 95%, and earned an Army Commendation Medal.

Built ML models to predict heart disease. Learned about non-parametric approaches and their tradeoffs.

Learned how facial recognition systems work through Principal Component Analysis in Linear Algebra.

First project to understand statistical modeling, predicting the S&P index from 15 macroeconomic indicators extracted from TradingView.

Made a website for my partner with Bugcat-Capoo GIFs.