Julio Ochoa

Payments program lead, now building software with AI. New York.

I make software with AI doing most of the building. This site shows the five products that came out of it, what broke along the way, and what I learned.

deterministic controls around nondeterministic intelligence

Eighteen years running programs that move money: tokenization, browser checkout, audit and controls. Since January 2026, one person operating an AI production line that plans, builds, verifies its own work and records what broke. Five products came out of it. This page is the record.

I write the ask. Scheduled AI workers build and test it. The numbers below are what that has produced.

Work goes in as a written ask and comes out as a merged pull request. Between the two, the line plans it, builds it, and proves it works before anyone sees it. Every change must pass its tests and an automated review; changes to deployment, to what the loop may do, or to its written rules wait for the operator to read them before they merge, and everything else merges once its checks pass and is read afterwards, with a revert path. What broke gets written down and feeds the next pass.

Computer engineer out of Venezuela. Online banking in Java before the cloud, then a decade of payments: tokenization, browser checkout. Now technology programs and controls: proving to auditors and regulators that a system does what it claims. Building with AI turned out to be the same job with a different kind of unreliability underneath. That is where the tagline comes from.

Perse is the half of an AI system that lives around the model, run like a program. Work enters as a Form with acceptance criteria naming a checkable artifact; scheduled AI workers shape, build, check and document it, and a worker plans against the repository and calls real tools. Failure is designed for: typed returns, parked premises, reaped sessions, and a liveness check that trusts evidence of a run over a worker's own report. Memory lives in Postgres, not the context window.

In regulated work the bar is not whether AI is impressive but whether its output can be reviewed, reproduced, and defended. A solo hackathon build, finished in 24 hours, turns business ideas into reusable AI artifacts and stops loudly when live inference fails. A curated-knowledge assistant discloses its confidence with every answer and says it does not know rather than guess. Every AI-assisted workflow carries subject-matter review before formal use.

What holds my attention: AI platforms and tooling that can be inspected and audited rather than merely trusted; governance built into the system instead of written next to it; and how a small team, or one person with a production line, ships work that holds up under partners, regulators, auditors, and production.

What I work on: AI platforms and tooling that can be inspected and audited rather than merely trusted, and governance built into the system instead of written next to it.