Writing
Essays on AI agents, markets, and payments.
How to Actually Evaluate an Agent
LLM evals grade answers; agents have permissions. A working framework for testing outcomes, process and safety before things break.
Theses
The bigger its permissions, the bigger the risk.
agent evals The stronger AI only needs to own the plan and the review.
fable 5’s playbook Security checks are scoring signals, not vetoes.
seraphina No card is perfect. If one claims to be, that’s an ad.
crypto cards, audited The root of every illegal profit stream is that the exchange is your counterparty.
the exchange playbook You will never out-calculate the bots on-chain. But you can out-calculate yourself.
don’t die on-chain
Work
Things made to run,
Syntaxis (opens in new tab)
A product studio in Palo Alto, building software you can hand a task to.
- A guess never becomes an action.
- A plan is never saved until you confirm it.
- Its long-term memory is a page you can read.
- A greeting never reaches the language model.
- Numbers are calculated, never made up.
- Every failure becomes a test.
Sila (opens in new tab)
A strength-training log for iPhone and Apple Watch. Log a set in one tap.
Examples it understands: 12 at 60, 8 x 135 lbs, same weight, rest 90. Anything else, like “a few at 60”, is not understood, and nothing is logged.
Type what you would say between sets, like 12 at 60 or rest 90, then press Enter. Nothing is logged unless the line is understood.
Seraphina
An autonomous trading agent that scores every token before it buys.
It scores 6 signals: buy/sell ratio, up to 30 points; smart money, up to 25 points; community boost, up to 15 points; candle trend, up to 15 points; liquidity, up to 10 points; contract security, up to 5 points. It buys at 60 of 100. Animated examples, not trades: example 1 scores 84, buy; example 2 scores 42, pass; example 3 scores 59, pass.
About
I build products and write about the systems behind them.
I work on product at Infini and run Syntaxis, a small studio in Palo Alto; Sila is our first app. Before Infini, I was a researcher at Sonic SVM.