Ross F. Williams
Computational Modeling · Generative AI · Full-Stack Projects
Building AI systems that model the real world: simulations, real-time analytics pipelines, interactive tools.
I hold a PhD in Industrial and Systems Engineering from Virginia Tech (2026), with research at the intersection of complex systems, artificial intelligence, and computational modeling.
I develop novel approaches that combine generative AI with traditional simulation methodologies. My work on epidemic modeling with generative agents was featured in The Daily Beast, demonstrating how LLM-powered agents can simulate realistic human behavior during disease outbreaks.
Before my PhD, I earned a B.S. in Mechanical Engineering from Virginia Tech (2021) and was part of the RoboGrinder robotics team that achieved 2nd place at the RoboMaster international competition.
My research focuses on developing computational methods that help us understand and predict the behavior of complex systems. I'm particularly interested in how generative AI can enhance traditional simulation approaches.
System Dynamics Review, 2024
We present a system dynamics bot that can automatically construct causal loop diagrams from textual descriptions, leveraging large language models to bridge the gap between natural language and formal system dynamics representations.
Read Paper →System Dynamics Review, 2024
This tutorial introduces generative agent-based modeling, a new paradigm that combines the flexibility of large language models with the rigor of agent-based simulation to model complex social systems.
Read Paper →arXiv preprint, 2023
We demonstrate how LLM-powered generative agents can simulate realistic human behavior during disease outbreaks, capturing complex social dynamics that traditional compartmental models miss.
Epidemics aren't just biology — they're nonlinear socio-technical systems driven by human behavior. People don't move through compartments like marbles through a machine. They notice rising risk, weigh trade-offs, change their minds, and act on incomplete information. Most classical models can't capture any of that.
The compartmental model that's powered epidemic forecasting since 1927: populations divided into Susceptible, Infected, and Recovered, governed by transmission rate β and recovery rate γ. Adjust the parameters to see how a homogeneous population responds.
But humans aren't compartments. They don't transition between states based on a fixed probability — they decide. An extrovert with kids and a service job behaves nothing like a risk-averse retiree, even at identical infection levels. Capturing that heterogeneity is exactly where mechanistic models break down.
So what happens if we replace the compartments with 100 LLM-powered agents, each with a unique persona, deciding day-by-day whether to shelter as risk rises?
100 generative agents with distinct personalities probe the same question across 40 infection levels: "Will you stay home today?" Each agent was asked 5 times per level to measure consistency — that's why the agent decisions chart shows confidence as 5/5 (unanimous) down to 3/5 (split). Scrub through the levels to watch the town empty out, or click any agent to see their reasoning.
Dataline Project, Virginia Tech
Led a $100,000 research project investigating how to leverage LLM agents for email marketing optimization and customer engagement strategies.
RoboGrinder: Robotics Team
Led Virginia Tech's largest robotics team (54 members) to 2nd place at the RoboMaster North American competition. Served as Sentry Sub-Team Lead for the international competition, where the team also achieved 2nd place.
BMW
Designed and optimized manufacturing processes for BMW's X series SUVs, addressing ergonomic and mechanical challenges on the assembly line.
MSE Journal Club, Virginia Tech
Founded and organized a platform for 15 Management Systems Engineering graduate students to present and discuss research papers.
ISE Department, Virginia Tech
Hosted orientation and welcome sessions for 60 new graduate students, helping them navigate academic and professional opportunities.
Computer vision pipeline that turns a ZED 2 stereo camera into a full pickleball analytics system. Detects the ball frame-by-frame with a fine-tuned YOLOv11 model, tracks it in 3D with an Extended Kalman Filter, segments rallies, and writes structured analytics to SQLite — all running on cloud GPUs via Modal.
A Progressive Web App that transforms Among Us game mechanics into a real-world phone-based multiplayer experience. Players join via codes, receive roles on their phones, and play in physical space.
A player rating and statistics tracking system for the social deduction board game Blood on the Clocktower. Tracks Elo ratings, win percentages, and game statistics with analytics dashboards and leaderboards.
A fully static analytics dashboard for Atlas Conquest, a competitive hex-grid deck-builder game. An automated daily pipeline pulls match data from AWS DynamoDB, aggregates stats across time periods and map types, and publishes interactive visualizations.
A geospatial agent-based model simulating a scavenger-hunt bar crawl along Atlanta's Beltline. Built on real-world locations and Google Maps walking times, groups navigate a 1.2-mile route searching for a hidden host — with realistic movement, information sharing, and process-of-elimination deduction animated live on an interactive map.
A native iOS astrology app built for a friend, featuring natal charts computed from scratch using Meeus astronomical algorithms and the Placidus house system. No external APIs — all planetary positions, transits, retrogrades, and synastry are calculated on-device.
During my time leading the RoboGrinder robotics team at Virginia Tech, we designed and built fully autonomous combat robots. This video showcases our robot in action at the RoboMaster competition, where we achieved 2nd place internationally.
Coverage of our research on using generative AI agents to model epidemic spread and human behavior during outbreaks.
Try the interactive epidemic model above to see compartmental dynamics in action, or read my papers on how generative agents can enhance these traditional approaches.
Try Interactive ModelI'm always interested in discussing potential jobs, research collaborations, speaking opportunities, or just chatting about complex systems and AI. Feel free to reach out!