#ai-assisted
Projects
Realistic Sensors
ActiveAI-assistedThree tools around one capture contract — an Unreal Engine 5 plugin that records synchronized, nanosecond-timestamped multi-modal sensor data (RGB, depth, segmentation, DVS, optical flow, IMU, LiDAR, thermal, cubemap); Forge, the Python analysis and fusion toolkit that reads those captures; and rsviewer, a browser viewer that puts every stream on one timeline.
TRANSFORM
ActiveAI-assistedA Modelica library for system-level dynamic simulation of nuclear and other thermal-hydraulic energy systems — fluid networks, heat transfer, heat exchangers, nuclear kinetics, media, and controls in one reconfigurable package. Started at ORNL in 2017, still moving, Apache-2.0.
Writing
One plugin, three repos
How RealisticSensors split into a UE5 capture plugin, a Python analysis toolkit (Forge), and a browser viewer (rsviewer) — the dates, the reasons, and what the capture and provider contracts between them buy.
Is my AI memory system obsolete? I sent 104 agents to find out
Six weeks after Soul System 2.0 shipped, I audited it against everything the ecosystem now does for free — built-in AI memory, Obsidian vaults, memory servers. The plumbing is commodity now. The conscience still isn't. And giving the record a graph found two typos in my own files within minutes.
Soul System 2.0: a conscience and a notebook for your AI coding sessions
Soul System 2.0 is out — one plain page of rules your AI coding agent reads, four commands with teeth, and a book for humans. The shape left standing after a measurement program tried to kill the rest of it.
I built the best opponent I could for my AI philosophy. We tied.
I ran the lean system against the best generic CLAUDE.md a careful session can write — real multi-session builds, scoring locked before anything ran. Every quality measure tied. The interesting part is why, and the two things that didn't.
I gave an AI a philosophy, then tried to prove it was useless
I ablated my own AI doctrine layer under equal compute — every apparent win re-run against length-matched filler. Most of it was theatre. What survived is memory of the things the model can't guess.
Ideas Are Cheap Now
We spent seventy years asking whether machines can be creative. We were arguing about the wrong word.
The same physics on Modelica + Dymola: reaching the full PWR assembly
The third episode of soul-test-netflow — Modelica 4.1 running in Dymola, the mature commercial simulator. The 17×17 PWR fuel assembly that Julia could only extrapolate, completed in 397 seconds end-to-end.
netflow: hand-rolling a conservation-network solver in Python
The first of three episodes in soul-test-netflow — a domain-agnostic "conservation on a network" solver in pure Python, with nuclear thermal-hydraulics as its first plugin.
Rebuilding netflow on ModelingToolkit: 14× less code, then a compile wall
The Julia / MTK episode of soul-test-netflow — same physics as the Python netflow, on a high-level symbolic modeling library. Matched node-by-node to 25 mK with 14× less code, and stopped short of a full reactor-assembly model when the compile step became the bottleneck.
Why this site exists
A home for the side projects piling up — and the stories behind them.