Hobby projects feeding evidence into the themes, each one a test of where AI's capabilities actually land in practice. Each keeps its own repo and site; this index links out and notes which themes it's feeding into.
A parametric pipeline for printing footprint-accurate Warhammer 40k 11th-edition tournament terrain on a Bambu X1C. Bases match the official footprint sizes exactly, host existing L-shaped buildings inside a slot, fill the rest with built-in scatter, and split into tiles with alignment features when they exceed the printer's build volume.
The connective tissue behind the game experiments. A repeatable chain from an empty project to a reviewed, asset-filled scene. create-godot-project stamps a template, the asset forges (audio-forge, texture-forge, AISCAD, TRELLIS) fill it with provenance-tracked art, a screenshot review harness proves each scene, and a web-export step ships it.
A vision-LLM agent that critiques plots, charts, and data displays for correctness and common misleading tricks (truncated y-axis, dual-axis sleight-of-hand, cherry-picked time windows, misleading area scaling, 3D distortion), then redraws an honest version from the data it can infer from the image.
Autonomous AI agent that watches a backyard chicken coop via Pi camera, cares for the flock, runs a weekly supplies budget, and publishes its observations as a static site (the "Coop Chronicle"). The origin experiment, and still the thesis anchor.
A workbench for designing and validating 3D-printed enclosures, brackets, and robot parts through a Claude-driven OpenSCAD + Godot pipeline. Components live in a shared inventory; assemblies are render-reviewed, collision-checked, and simulated before printing.
A trio of harnesses that point a local model at real open problems and demand verifiable results. erdos-at-home writes and runs finite Python searches against open Erdős problems; physics-bridge proposes cross-theory bridges behind a dimensional-homogeneity gate; swarm-topology benchmarks which agent communication shape clears a task queue for the least compute.
Less a game than an AI bake-off. Three architecturally different attempts at the same VR scene (dual-perspective mech combat, a third-person miniature on your hand plus a first-person cockpit) are built in parallel, auto-gated, and vision-judged against reference images until one clears a weighted score.
A single persistent "Writer" agent that owns and grows a body of creative work, chooses its own direction, and hands off to its next self at the end of each session. Left with an optional seed (or nothing at all), it self-paces, reconstructs its own context, and journals the decisions it makes between pieces.
Natural-language to print-ready wargaming terrain. The user describes a building ("L-shaped townhall, three floors, gothic, balcony on the second floor"); the system selects pieces from modular STL kits, generates a layout, gets human approval, and produces split-for-printer STLs with alignment features preserved.
A SteamVR overlay plus an always-on voice loop that turns a headset into a hands-free control surface for Claude Code. Stay in VR iterating on a scene, speak feedback, say "go ahead", and it hands a clean brief to an agent rooted in the right repo. A world-locked panel shows live fleet status (context window used, busy state, progress).
A room-scale VR "living orrery". Stand over a solar system, grab and scale planets, drag orbits, warp time, then descend to any body's surface where its moons wheel overhead in real orbital geometry. A live Newtonian n-body sim (velocity-Verlet, Plummer softening) keeps motion continuous across the god-view and the surface view on one shared clock.
A PCVR sword fighting game where opponents are "ghosts" replaying recorded human motion. Records 60Hz pose data, tags outcomes, evolves a population of sequences via Elo and tournament selection, and serves the strongest as the next opponent.
A long-form story generation system built around an MCP server for story arc, state, characters, locations, and chapter review. Chapters are produced sequentially by Task subagents that read only the prior chapter plus shared state, targeting novel-length output without a single mega-context.
More projects are candidates. A project becomes an experiment once it's produced a real artifact and at least one finding worth a post.