Building world models where humans, robots and AI agents learn, play, and evolve together.
I spent many years making AAA games, then moved to simulation for defense and robotics. My environments train and test robots and AI agents before they act in the real world, so they protect people and stay aligned with the humans they serve.
Six projects with public code, live demos or video. Each project page covers the problem, the design and the results.
A self-hosted platform for AI agents that work with physical systems. Agents read sensors and video, fly simulated UAVs in Isaac Sim with PX4, control SUMO traffic, and record every run for replay.
An agent flew 9.227 km over Manhattan in 13 min 10 s: four waypoints, zero collisions.
Procedural 3D worlds for training vehicle agents with PPO. Twelve vehicle types across ground, sea, air and underwater. The simulation, sensors and trained policies also run in the browser.
A 49.5-million-step reliability test kept a candidate policy that failed its gates out of the public demo.
I developed the simulator shown in Shield AI’s Hivemind fixed-wing demo. Hivemind is Shield AI’s autonomy software for military and commercial aircraft.
The demo video is from 2021. Shield AI interviewed me as a dynamical systems engineer in 2018.
An execution model for simulated worlds. A Lean 4 kernel checks every state change, and each run can be replayed. Agents, people and planners propose actions; the kernel accepts or rejects each one.
Seven interactive showcases. Foundry: Night Shift steps through 95 checked snapshots.
A Rust engine that models a problem as accounts, assets and exchanges. Search, optimization and reinforcement learning propose exchanges; the engine validates each one and keeps a replayable trace.
Fourteen problems share one service contract across the CLI, HTTP, MCP and the browser Studio.
An open-source simulation engine and AI platform for training perception models on synthetic data and testing robots in simulation.
Unreal Engine 4.24 with NVIDIA PhysX 4.0, plus plugins for synthetic data, GStreamer pipelines and dynamical systems.
Short videos of current experiments. Hover over a card to preview it; select it to watch.
One learned policy across every test scene, including missing limbs, slopes and moving platforms. Final evaluation: walking 36/36, static balance 136/144, moving supports 24/24.
Simulated rovers share data over a Reticulum mesh network with heavy packet loss.
Flight control for a simulated fly, driven through the MaleCNS connectome.
A coding agent and reinforcement learning in one development loop for the AlienWars game.
A small encoder model picks Doom actions at 43.7 decisions per second, 13.9 ms each, on an M3 Max.
A quadruped policy trained across 4,096 parallel worlds with 400 hours of simulated experience.
A quadruped with leg-mounted floats walks into water, floats and walks out. Built in 40 minutes from a design posted on LinkedIn.
An agent implemented the SkyJEPA paper and trained a UAV controller with 26–34% lower worst-case trajectory error than a geometric controller.
A physics-informed neural operator forecasts wind for package delivery in MuJoCo. It delivered in 5 of 6 held-out wind cases with 77% lower tracking error.
Quadcopter dynamics and depth sensing split across processes and recorded in Rerun.
Text and image prompts drive an agent that generates 3D characters procedurally.
NPC agents with individual goals that produce group behavior.
Veoveo, AlienWars Gym, Maquina and Axionomy, plus the experiments in the lab above.
Principal AI and Simulation Engineer (contract). AI and simulation infrastructure, harness, training, synthetics and building world models where humans, robots and AI agents learn, play and evolve together. Developed the simulator shown in the Hivemind fixed-wing demo.
Founder & CTO. Building scalable simulation infrastructure solutions and immersive experiences centered on human-robot-AI interaction. Designing and implementing multi-agent frameworks and behavior systems.
Lead Simulation and AI Engineer. Leading development of an open source physical AI platform for sim-to-real transfer and digital twin applications, and real-time synthetic data pipelines for training embodied AI models.
Consulting: sensor characterization and visualization tools, Unreal Engine in development operations, and autonomous swarm simulation architecture for DARPA and DoD programs.
Real-time deep learning inference in a VR music MMO using TensorFlow, GStreamer and Unreal Engine 4; physics-based VR gameplay and full-body VR networking.
Designed and built a robot-crafting sandbox game with physics-simulated robots (Bullet Physics).
Gameplay, physics and networking on console titles, including Star Wars: The Force Unleashed. Simulation tools for real-time effects and previsualization on several films.
And
yet
it moves!
Proposes modeling digital twins with three primitives: objects, operations and machines. The idea became the Maquina project.
Argues for measuring how far a virtual world is from equilibrium instead of designing for perfect balance.
Describes Situation Design: areas of activity with a radius of influence, role slots and growth rules, placed above agents and encounters.
Proposes designing worlds from meanings such as “resting place” and storing memory in places and objects as well as in NPCs.
Where a small encoder model fits in games: classifying player behavior, game state and procedural content.
Actor model, behavior trees, memory systems and language models for agents in virtual worlds.
I started in console games, writing gameplay, physics and networking code at LucasArts and other studios across the industry. Since 2016 I have worked on simulation for robotics, autonomy and defense, and on the perception and learning systems trained in those simulations.
I prefer small sets of well-defined parts: typed state, explicit rules, and runs that can be replayed. I write Rust by choice, C and CUDA where performance requires it, and Python for training and tooling.
“With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.”
| Rust | Veoveo, Axionomy |
|---|---|
| C, CUDA | AlienWars Gym |
| Lean 4 | Maquina |
| PPO, PufferLib | AlienWars Gym |
| MuJoCo | lab experiments |
| Isaac Sim, PX4 | Veoveo |
| Unreal Engine | Simbotic, Shield AI |
| Unity | games archive |
| GStreamer, LibTorch | Simbotic |
| WebAssembly | AlienWars Gym, Axionomy |
| Model Context Protocol | Veoveo, Axionomy |
| Rerun | Veoveo |
