The RealEyes.ai project is an award winning deepfake detection system designed to combat the growing threat of AI-generated misinformation. With deepfake-related fraud expected to cause billions in losses and contribute to cybercrime, harassment, and global disinformation, RealEyes provides a user-friendly, transparent, and accessible solution for identifying synthetic media.

RealEyes employs a multi-model machine learning approach to detect deepfakes with high accuracy. It integrates a Chrome extension and a web-based image upload tool, allowing users to analyze images seamlessly in real time. The system offers features such as probability ratings, analysis details, and a feedback system to improve detection capabilities. Unlike proprietary algorithms, RealEyes maintains transparency and accountability by being open-source.

I did all of the development of the project in Node.js, HTML, Javascript, PHP, css, and deployed in AWS using AWS Cloud Formation, Lambda, DynamoDB, API Gateway, SageMaker for serverless inference.

My final cryptography project provides open-source Wolfram Mathematica scripts for visualizing elliptic curves over finite fields by mapping them onto a 3D torus. This project builds on the work of Pantůček, who demonstrated the only depiction of a Weierstrass curve mapped over a torus that I could find.

This project allows one to see how the elliptic curve wraps in 3D space, intersecting with finite fields at rational points ⬤.  One can glean insights into the distribution of rational points and their relationship as the field size increases.

Visualizations include comparisons across different types of curves, like Weierstrass and Montgomery.

Find it on GitHub here!

In this project, we explored the prevalence of Personally Identifiable Information (PII) and Quasi-Identifiers (QI) across over 291,000 datasets hosted on data.gov. Our research identified challenges in relying solely on metadata and demonstrated the value of analyzing underlying data to accurately estimate privacy risks. We leveraged the open-source tool DataProfiler, created by CapitalOne, to provide insights into the prevalence of PII and QI in government hosted datasets, while highlighting opportunities for future work, such as developing automated machine learning models to identify Quasi-Identifiers.

Find it on GitHub here!

I developed QuasID, an open-source machine learning model that automatically detects Quasi-Identifiers (QIs) in datasets – a component for privacy engineering and data protection. The project addresses a gap in open-source privacy tools, as while solutions exist for detecting Personally Identifiable Information (PII) (such as the DataProfiler by CapitalOne), there weren’t accessible tools specifically for QI detection.

The solution leverages Meta’s Llama 3-8B model (8 billion parameters) with Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA) techniques. To make the model practical for real-world use on consumer hardware, I implemented memory optimization strategies including 4-bit quantization using the bitsandbytes library. The model was trained on manually labeled datasets from UCI’s Machine Learning Repository, achieving decent accuracy metrics on the test set. The entire implementation is documented in a Jupyter notebook with step-by-step instructions for reproducibility. Key technical achievements include successful fine-tuning of a large language model on consumer-grade hardware, effective memory management through quantization, and creation of a specialized training dataset for QI detection.

The Foster Care Legal Network runs a program called Simply Friends, which aims to enrich the lives of children in foster care by introducing meaningful and healthy relationships and experiences. As a board member, I was tasked with creating the simplyfriends.org website.

I used WordPress Divi to build the site, which is also used for this site!

The Simply Friends program has expanded to several foster group homes in the Los Angeles area.

We moved during the pandemic and decided to try and remodel our home despite part shortages and supply chain issues. Delays forced us to live out of our garage for several months, but the outcome was worth it.

My college roommate and architect friend, Jake Hofeling, designed the layout of the kitchen and living room. Doug and Adam Chamorro from Hardwood Habitat did the built-ins and island, which came out beautifully.

My dad originally bought this 1955 GMC Pickup Truck for $800 when I was 6 years old. I asked him if we could restore it. Little did I know how much work that would be. We patched rusted holes in the step-sides and quarter panels, converted the front brakes from drum to disk, and ran new brake lines. We picked out white oak planks, and cut/planed them to size. We applied POR-15 to the frame and bed running strips. The engine was rebuilt. Then I went to college, and we decided to sell it to a close family friend.

They completed it from the frame up, but unfortunately passed away while re-assembling. My dad and I bought the truck back, and we did the final assembly. It didn’t fare well in the traffic of Los Angeles, so we sold it to a nice family in Colorado.