How I work

Find the leak before choosing the tool

Every system I have shipped started as a conversation with the people doing the work: a lawyer running a title search, a team lead listening to calls, an engineer staring at a cloud bill. The question is never what could AI do here. It is where hours or money are leaving the building, and how much. Until the leak has a size, there is nothing to design.

Most of the leaks turned out to be boring: forms, portals, recordings nobody opens, bills nobody can read. That is the point. The dramatic use cases are rarely the expensive ones.

One hardened stack for every system

The first system took the longest, because it carried the cost of the base: authentication with the company's single sign-on, roles, a job queue, monitoring, a security review. Every system since has started from that base. Nothing is built twice, the security posture is the same everywhere, and a new system is a question of weeks.

The stack itself is unremarkable on purpose. A typed backend, a typed frontend, a queue, a database. Boring choices are the ones a new hire can read at three in the morning.

A measurement layer in every build

Each system carries its own view of what it did: how many cases, how many hours, how much money, for whom. That view is not a report I write afterwards. It is part of the product, visible to the business owner on the day they start using it.

The habit changes two things. The owner can repeat the number in rooms I am not in. And the view itself often becomes a second product, because a record of what people ask a system is a map of where the business is struggling.

Open models, tuned for the job

In a regulated business the data should not leave, and a bill that grows per call is a bill that grows with success. So wherever an open model, repurposed and tuned for the task, is good enough, it runs on infrastructure the company controls. The cost becomes a server, not a meter.

Good enough is measured, not assumed. Each model is tested against the vendor or the human it replaces, on the same cases, before anything is switched off.

How a project starts

I spend part of every week with a business team, especially in the parts of the company technology has not reached: branch operations, disbursement, servicing, audit. A candidate becomes a one-page scoping note with four lines:

• The leak, in hours or money.
• The smallest system that would stop it.
• The owner who will use it every day.
• The number we will measure to know it worked.

Nothing is built without an owner and a number. Anything still without an owner a month into a pilot is stopped. That rule has killed more ideas than any technical limit, and the systems that survived it are the ones in production.