I build games, simulations, AI systems, and developer tools.
I work primarily in TypeScript/JavaScript and Python across agent systems, services, automation, browser applications, real-time 3D, and GPU-driven simulation. Agent work is language-flexible too: the Strands Agents SDK is my go-to harness, whether the surrounding system is TypeScript or Python.
I choose languages and frameworks to fit the system rather than treating the stack as an identity. My recent public work happens to lean TypeScript, while much of my production work lives in company repositories. I am also building fluency in Rust through smaller experiments.
Open MK OS - my interactive portfolio and an executable index of the wider practice. It is also a project in its own right: a hand-rolled TypeScript browser system with its own window manager, virtual filesystem, simulations, and WebGPU applications.
sheepdogsim.com · source
Browser-based 3D sheep herding game. Solo leaderboard modes, 2-4 player online co-op and competitive rooms, split-screen, a sandbox editor, full mobile + gamepad support. Three.js with WebGL fallback and progressive WebGPU, React 19, Cloudflare Workers + Durable Objects. Fully open source (AGPL-3.0) - a complete shipped multiplayer game you can read end-to-end.
terror-in-the-jungle.pages.dev · source
Combined-arms Vietnam-theater FPS/RTS sandbox // infantry, ground vehicles, helicopters, and fixed-wing aircraft in one runtime, architected for ~3,000 NPCs on real A Shau Valley DEM terrain. WebGPU/TSL, 314k lines, 5,900+ tests.
objekt62.com · closed source
Multiplayer salvage-horror in the browser. Authoritative-server netcode on Cloudflare Durable Objects with client-side prediction and reconciliation, deterministic CPU/GPU terrain parity. The code is private IP - the game is live.
kilnstudio.tools · closed source
Text-to-3D GLB generation with a pairwise model-ranking arena, deployed on AWS - Bedrock AgentCore Runtime, ECS Fargate, DynamoDB, Cognito, CDK-managed infra. The current endpoint of three years of AI asset-pipeline work.
| sds-gpu-boids | WebGPU compute-shader flocking sandbox. 100k+ agents, prefix-scan spatial grid, CPU reference oracle validating the GPU pipeline. The lab behind Sheep Dog Simulator. Live experimental build |
| goldberg-planet | Demo concept: a walkable, flyable, mineable planet whose surface is a true Goldberg polyhedron, topology derived from scratch. Live demo |
| mycelium-v2 | Multi-agent orchestration for CLI coding agents. Deliberately archived - a frozen snapshot of what agent orchestration looked like in early 2026, kept as reference for the next rebuild. |
| falling-sand-alchemy | Zen falling-sand game about transmutation. Pure cellular automaton in TypeScript drawn through a WebGL2 bloom pipeline. Play it |
| wavy · closed source | WebGPU ocean sandbox on Three.js TSL. A CPU/GPU twin: the same Gerstner wave math drives the GPU water surface and CPU-side buoyancy for drivable boats, with compute-shader foam and wake and dual above/below-water rendering. The lab behind the ocean work mentioned below. Live build |
GPT-3 experiments (2023) → local model pipelines and vanilla-JS games (2024) → WebGPU engines, shipped multiplayer games, and production AI systems (2025-26).
The early public repos are still up: hand-rolled RAG with raw embeddings, radar (proto tool-calling and context recursion before that was standard tooling), 2DGAG (a local Stable Diffusion asset pipeline), zoo-mcp, plus a WebGPU ocean sim, a whale sim with sonar-as-vision, a PSX-style survival horror, and early experiments running coding-agent swarms unsupervised. Most of that stays private - old prototypes and stepping stones.
The heaviest AI work lives in the company repos of a previous role: a year of building enterprise AI pipelines - a dozen MCP servers (half deployed to Bedrock AgentCore), agent skills, case triage and resolution pipelines, classifiers, bug-hunting agents - plus a full-stack dev portal and a support portal, shipped solo. kilnstudio.tools exists partly to show that same stack and scale in public, built solo.
Before all of this I worked as a senior blockchain systems analyst. I intend to come back around to crypto eventually - there are systems I designed back then that the tooling couldn't build yet.
If there is one through-line, it's simulation. Boid flocking from CPU prototypes to 100k+ agents in WebGPU compute (sds → sds-gpu-boids), cellular automata (falling-sand-alchemy), a streaming voxel planet on real polyhedron topology (goldberg-planet), combat simulation for thousands of NPCs on real USGS terrain (terror-in-the-jungle), plus private work on Gerstner-wave oceans, whale kinematics with sonar-as-vision, and hand-rolled flight physics. Games are how I ship simulations people can touch.
Some repos here are deliberately frozen rather than maintained. Model generations move fast; a working orchestrator from six months ago is worth more as an artifact than as a codebase to drag forward.
Recurring habits: CPU/GPU parity (terrain math, wave math, and boid logic that run bit-identically on both), hand-rolling core systems instead of reaching for engines, and shipping live URLs - a project isn't done until someone can play it.