William Keffer
Computer Science × Mathematics × Machine Learning
Bio
William is a sophomore pursuing a BS in Computer Science and Mathematics at UNC Chapel Hill. His research includes adversarial machine learning, vision-language model security, and satellite-economic analysis. His primary interests now are machine learning and the intersections of analysis and abstract algebra with computational methods. He loves talking about all things related to these fields. In his free time William enjoys designing redstone contraptions in technical Minecraft, tending his Stardew Valley farm, and listening to Pitcher 56 and Aphex Twin.
I built this site to reflect some of the design styles and aesthetics that resonate with me. Inspired by the abstract work of Aletiune, nine rendering layers stack from back to front forming a highly chaotic and variable scene. This portfolio runs on Next.js 16 with TypeScript, Tailwind CSS, and Zustand, rendering a multi-layered composition of canvas-drawn chrome strokes, drifting sky fragments, particle dust, and glitch artifacts at 60fps. A date-derived seed feeds a Mulberry32 pseudo-random number generator each day, deterministically driving the placement, curvature, and styling of every visual element so the composition stays consistent within a single day. When the date changes, the seed advances and the entire visual landscape regenerates, ensuring the site looks different every day while maintaining its cohesive aesthetic.
Experience
Blockd
Blockd is an enterprise-grade interview security platform that combines a custom Chromium browser with AI-powered anti-cheating detection to preserve the integrity of remote interviews. The platform deploys a suite of real-time monitoring tools, including eye tracking via MediaPipe, multi-LLM answer semantic analysis, mode collapse screening, process detection for screen recording software, VM detection, and window focus tracking to identify candidates who may be using unauthorized assistance during live interviews. On the interviewer side, a React-based web dashboard provides reports on security events, gaze heatmaps, AI detection confidence scores, and live video feeds in real time, giving hiring teams full transparency into each session. Built on a microservices architecture with eight backend services, WebRTC video streaming, WebSocket real-time communication, and a Kubernetes-ready infrastructure.
Experimental Engineering Lab
The Experimental Engineering Lab (EEL) is an undergraduate research lab at UNC dedicated to hands-on engineering projects and interdisciplinary collaboration. I have been responsible for designing, developing, and maintaining the site, managing its deployment on Red Hat OpenShift through an automated CI/CD pipeline.
Duke University, Dept. of Anesthesiology
Conducted a meta-analysis with anesthesiologists on the effects of anesthesia on perioperative sleep outcomes, applying statistical modeling and systematic review methods across clinical studies. Gained experience in data synthesis, cross-disciplinary collaboration, and clear communication of research findings.
Projects
Brick Studio
A desktop application that lets anyone describe a Roblox game in plain English and receive a complete, publish-ready game file in return. The tool transforms natural language prompts into fully functional Roblox places containing server and client scripts, user interfaces, physics, lighting, and data persistence. Under the hood, a TypeScript MCP server orchestrates nine specialized AI agents powered by Google Gemini through a recursive function-calling loop, while a Rust serializer built on the rbx-dom ecosystem produces the binary .rbxl output. The Electron + React frontend provides an IDE-like experience with a conversational chat panel, live code preview, build progress tracking, and one-click export.
Argo
An adversarial robustness research tool targeting the most underexplored attack surface in modern AI: the projector layer connecting vision encoders to large language models. This framework demonstrates that the thin projector bridge in VLMs like GPT-4V, LLaVA, and BLIP-2 is disproportionately sensitive to perturbation, often exhibiting higher Lipschitz constants than the vision encoder it wraps. Implements three complementary attack strategies (feature divergence, cross-modal misalignment, and visual token corruption) using PGD optimization with SPSA gradient estimation, all operating within imperceptible perturbation budgets.
Aperture
A satellite-economic ML pipeline for real estate investment intelligence, built at Carolina Data Challenge 2025. Processes NASA HLS satellite imagery with computer vision to detect construction, fuses results with FRED/BLS/Census economic indicators, and scores development opportunities using PyTorch and XGBoost ensembles. Deployed on AWS (SageMaker, Lambda, PostGIS, S3) with Dask-based distributed processing and containerized inference.
Extracurriculars
Natural Sciences Computing Club
NSCC bridges the gap between traditional science education and the computational skills essential to modern research. We create computational models for natural sciences research, and our goal is to build a community where scientists confidently use code to advance their fields, cultivating the next generation of computationally fluent researchers ready to work at the intersection of data and discovery.
Visit site →UNC Student Government
Rebuilt the Executive Branch website from scratch, migrating off WordPress to a custom application on Red Hat OpenShift, and manage the full deployment pipeline including DNS, SSL/TLS, and cloud infrastructure.
Visit site →Undergraduate Research Conference at UNC
Designed the conference logo, built the event website, and manage the conference's full digital presence. The site serves as the primary platform for applicant outreach, event information, and sponsor engagement for a one-day research showcase in the Campus Union's Great Hall.
Visit site →Let's connect
Open to collaborations, research opportunities, and conversations about the future of computational design. Reach out — the best projects start with a message.
Designed & built with Next.js, TypeScript, and generative algorithms.
The compositions are generative and change daily.