AI · How we build
Six rules we run our own shop on.
These aren't aspirations. They're the rules we work under every day, and anything we build for you gets the same ones.
A person signs off
Anything a customer will see, anything with money in it, and anything you can't undo waits for a human okay. The sign-off is a step in the workflow, not a line in a policy.
Privacy by default
For sensitive data we run open models on hardware you control, so it never leaves. Hosted models only where that's fine.
We check the output
Anything a model produced gets checked against the source data before anyone acts on it. "It worked" isn't evidence. Reading the result back is.
Tested before launch
We run a test set, including the prompts people use to break things, before a workflow goes live and again after every change.
Logged, with an off switch
Every action gets logged. There's an off switch, and we write down how to undo things before we roll them out.
The right model for the job
We use several models and pick by the task, the cost and how sensitive the data is. We keep the vendor names off this page on purpose.
What privacy looks like in practice
The Funnel Check on this site is a small example.
Right now it's a set of rules. If we later have an AI model write the readout, that model will run on our own hardware, so your answers never go to an outside service. And if we run that experiment, we'll publish how it did.
Want these rules on your workflows?
The assessment turns up three to five candidates, ranked. You pick which one becomes a pilot, or none.
Fixed fee, quoted after a 30-minute call. We reply within one business day.