Robotics
Whole systems and, more often, the parts inside them: actuators, harmonic drives, sensors, the supply chains that make them.
01 — The desk
A trading operation where every trader is code
There is no discretionary layer. No moment where I look at a chart and decide. The desk runs a fixed sequence: candidates are found, a thesis is built and attacked, risk sizes it or kills it, execution works the order, and every decision is written down and read back later.
My only input is the mandate. When something goes wrong, I don't take the wheel — I change the rule that allowed it and let the desk run again. That constraint is the whole design. A decision made in the moment is worse than a decision made by a system, and the fastest way to prove you believe that is to remove your own ability to intervene.
02 — The traders
Six mandates, one veto
People call these AI agents. That is accurate and it is also the most worn-out phrase on the internet right now, so here is the version that means something: each one is a bounded piece of code with a written mandate, a defined scope of authority, and a record of every call it has made.
| Trader | Mandate | Authority |
|---|---|---|
| scout | Scans the instrument universe on formal criteria and produces candidates. Knows nothing about what is fashionable. | proposes |
| analyst | Builds the thesis and then spends most of its budget trying to break it. A candidate survives only if the attack fails. | proposes |
| risk | Owns sizing, exposure and stop conditions. Reads every proposal against the whole book, not the single idea. | full veto |
| execution | Routing, timing, slippage, liquidity. Will refuse a fill it cannot get on acceptable terms. | partial veto |
| auditor | Reconstructs every decision after the fact and holds the desk's memory. It is the only trader allowed to say I was wrong. | reports |
| router | Moves work between the others, resolves deadlocks, and escalates to me when two mandates genuinely conflict. | escalates |
The hard problems here are not machine-learning problems. They are management problems: scope of authority, escalation, when to trust a call you disagree with, and when to shut something down.
That is the part I did not expect. I have run the same arguments with software that I would have run with people — except the software does exactly what its mandate says, which means every failure traces back to something I wrote.
03 — What the screen buys
Components, not themes
The desk decides where capital goes as well. It screens for what the next decade physically runs on, and it looks at components rather than stories — the things that have to exist regardless of which company wins.
motion
autonomy
Drones & autonomous systems
Air, ground and marine platforms, plus the navigation and control stacks that make them usable without a person.
compute
Memory & silicon
Memory, interconnect, packaging and the unglamorous parts of the compute stack that constrain everything above them.
substrate
Data centers
The buildings, cooling and land that turn silicon into capacity, and the constraints that decide where they can be built at all.
frontier
Quantum computing
Hardware, error correction and the control electronics around it — held to the same screen as everything else, on a longer clock.
power
Energy, nuclear included
None of the above runs without power. Generation, grid, and the reactor programs now being built to serve compute directly.
New listings get the same treatment as anything else. A company that went public last week and one that has been trading for a decade enter the screen through the same door, and the screen does not know which is which. Candidates come out of it — not out of conviction.
When I write about a name here, I hold it. That will always be stated.
04 — Writing
The log
Everything the desk does that can be published, is. Three recurring formats:
room logWeekly. What happened on the desk: which mandate changed, where two traders disagreed, what broke. Mechanism, never returns.
the screenMonthly. A public run of the screen across compute, power and autonomy — the criteria, and what came out of them.
vetoOccasional. A trade risk killed, and whether it was right. These are the ones worth reading.
Entries go here as they are published. Nothing on this site is a recommendation, and no performance figures appear anywhere on it — see the colophon.
05 — Off the desk
The zeroth mandate
I race triathlon. One full Ironman, more than eight half distances, and an Olympic in Boston next. It is not a hobby I keep beside the work; it is where I learned the rule the work is built on.
A race is not won by effort on the day. It is won by a protocol written twenty weeks earlier and executed on mornings when nothing about it felt reasonable. How I feel on a given Tuesday does not get a vote, in exactly the way my opinion of a chart does not get a vote. When the protocol fails, the answer is to fix the protocol, not to override it once and feel strong about it.
So I hold myself to the same structure I hold the desk to: a written plan, measured load, a stop condition, and an honest review afterward. I break it about as often as the desk does, which is the useful part.
I am also learning to fly, and looking at a two-seat aircraft for short hops. It is the same discipline in a less forgiving room. The checklist exists because judgment in the moment is the least reliable instrument on board, and the go/no-go minimums are set on the ground — before there is any pressure to bend them.
06 — Now
What I'm working on
deskRebuilding how the traders talk to each other — the interaction protocol between them, and how a decision actually gets reached when two mandates disagree. Everything downstream depends on that layer, and it is the part I understand least well.
trainingOlympic distance, Boston.
learningTo fly. Looking at two-seat aircraft for short hops.
locationBoston. If you're here and working on any of this, say so.
updatedAugust 2026
If you are building something in this direction — autonomous execution, frontier hardware, or the boring infrastructure underneath both — write to me at [email protected]. I answer everything in the first day.
07 — Colophon
What this is, and what it isn't
- I trade my own capital. I do not manage outside money and I am not a registered investment adviser.
- Nothing here is investment advice, a recommendation, or an offer of any kind. It is a record of how one system works.
- I do not sell signals, subscriptions, courses or mentorship, and I do not answer "what should I buy".
- I do not take payment to write about any company. When I name one I hold, I say so in the same sentence.
- No performance figures appear on this site. Any backtest shown is hypothetical by definition and labeled as such.
Built and written by me. [email protected]