Operating Notes
Company-Building Under Real Constraints
Decisions about customer concentration, automation, services, enterprise risk, and pivots.
I work on AI agents and the systems around them.
I’ve spent most of my career building software for payments, healthcare, and travel.
These days I work on AI agents. I study where they help, where they fail, and what has to sit around the model.
I build prototypes, run agents in my own workflows, and write about what breaks. Sometimes the answer is a better model. Often it is clearer state, a verification step, or ordinary code.
Current work
TraceGuard is a research project on checking an agent’s work against the systems it was meant to change. The prototype uses synthetic travel workflows and injected failures.
Deterministic Offload replays an agent’s logged decisions against the operator’s own corrections to find the calls a model never needed. In 631 of my message-routing decisions, six written rules represented my policy better than the model did. Code makes those calls now, and the method is in a short paper: When Can You Delete an LLM Call? (PDF).
PillWatch is a small computer-vision system I built with four high school students. It forced us to think carefully about small datasets and system-level error.
Writing on AI
-
The LLM call you can delete
631 replayed decisions, six written rules, zero disagreements. The model I was paying to route my email lost to the policy file it was reading.
-
An AI task is not done until the world has changed
I moved my working life onto AI agents and logged every failure. Each one traced to the same cause: the to-do item is the wrong primitive for a machine.
-
PillWatch: 99 percent accurate, wrong every fifth box
Four high schoolers, a $60 Raspberry Pi, and the compounding math that turns 99 percent per cell into an error every fifth box.
-
1,029 cells: what small, honest data teaches about AI
The PillWatch team labels its own data, splits it by photograph, and hunts its rare cases deliberately. That discipline taught more than model tuning did.
Routespring
Travel software for the cases that do not fit the form.
At Routespring, I am testing where voice and language-model agents help with crew travel and disruption handling. The hard part is not generating an answer. It is knowing whether the work is complete.
Before Routespring
I helped build Pine Labs, GlobalLogic, hCentive, and Gallop.ai. The company history has the dates, roles, and sources.
Up Robotics
Co-founder and advisor
A hands-on robotics and engineering program for young students. I help with strategy, technology, and the occasional problem that needs more adults than children.
Off Duty
Away from work: robotics, over-built home infrastructure, and ranked lists of films and shows I keep thinking about.
Contact
For AI, airline, travel, founder, or Up Robotics conversations, use the contact page.