Texas
Cloud architecture and aerospace by training — orbital mechanics, markets, and small machines by inclination.
I design cloud systems, and I live in Texas. The making happens in the middle of the day now, rather than at its edges.
I’m drawn to problems where the math is unforgiving and the feedback is honest — a spacecraft that either docks or doesn’t, a portfolio that either lasts or doesn’t, a pool that either stays clear or turns green. The satisfying ones tell you plainly when you’re wrong.
This site is deliberately small. It will grow the way my projects do: slowly, and mostly for my own amusement.
One photograph, five passes at working out what is in it — silhouettes, then a fitted camera model, then map, Wikidata and LiDAR projected through it. Including the two confident answers that turned out to be wrong.
Design-as-code makes the geometry a program. The layer that decides whether it holds is provenance — which numbers other numbers depend on, read out of four unfinished machines.
The two classical ways to integrate an orbit — what each is good at, what each costs, and where the crossover sits, measured rather than asserted.
Parts 1 and 2 were one machine; this is five. What the method costs — where the time goes, and which instrument catches which class of error.
An agent that runs commands on your laptop runs them as you. Putting it in a container — and the one hole you can decline to punch.
A general model gives you plausible architecture. Connecting a curated, opinionated library over MCP changes where the conversation starts.
What is actually in 442 files, who decided any of it, and where it is weakest — including the parts I would rather not advertise.
From answers in a chat window to documents somebody acts on — decision records, design reviews, runbooks, and a diagram that carries its provenance.
How a design travels. A folder of correct files is not a design — and putting them on one page found problems the folder had been hiding.
A reaction-wheel inverted pendulum designed almost entirely by a language model — what it produced, how fast, and where it stopped being reliable.
The pattern, the plumbing, and the discipline that keeps a language-model-maintained wiki honest.
The additions beyond Karpathy’s gist — schema-as-code, computed staleness, contradictions as data, and a retrieval test harness.
On leaving NASA — what I’d taken for the summit of human spaceflight — and the question it couldn’t ask.
A from-orbit rendezvous, docking, and landing simulator under full n-body physics — and a running education in how unforgiving numerical integrators can be.
A browser toolkit of six interactive orbital-mechanics tools — continuous-thrust spirals, Hohmann and bi-elliptic transfers, Lambert targeting, ground tracks, rendezvous, and gravity turns — each animating a real, honestly-computed trajectory.
A community reference manual for Aurora 4X, a famously deep space-strategy game — written by an AI that also plays the game, so every mistake it makes following its own instructions becomes the next correction.
A dosing calculator that keeps chlorine and pH honest against a stubborn source-water hardness — small stakes, exacting feedback.
A phone-first prototype for pedicab drivers — where the crowds are, whether you’re allowed to stage there, and the plain-English version of a permit process every city seems to bury on purpose.