Blog
Writing on backend, architecture and product development.
I built a system that decides whether an industrial migration can be quoted at a fixed price. Its suites were green for days over data no user could have produced. Four ways of passing without having checked anything.
I built a local meter to find out what AI agents cost me. The four traps I hit were not AI traps: they were accounting traps. Cumulative totals that must not be summed, cache counted twice, and filters cutting in the wrong place.
A transactional backend does not fail the day it goes down. It fails the day it works twice. Idempotency, double-entry and consistency when one operation crosses three systems that share no transaction.
The script finished green, the file arrived on time, and every row said "unknown". No exception, no red log. The worst failure is the one that does not complain.
No puedes pedirle a un negocio que deje de cobrar un fin de semana mientras reescribes su núcleo. Shadow-first: el camino nuevo corre en paralelo, sin tocar nada, hasta que demuestra que coincide.
Empecé con la robótica como descanso del trabajo. Resultó ser el mismo problema con otra ropa: un motor que ya giró y un pago que ya salió tienen exactamente el mismo botón de deshacer, que es ninguno.
The module was built and tested. All that was left was turning it on. The question that was not in the ticket: what moves on the first run?
Un helper decidió que el punto de "1.500" era separador de miles. El error se multiplicó por mil en un camino de dinero. Y había un test verde que garantizaba ese comportamiento.
I build payment systems with AI agents every day. What it taught me most is not what they can do — it is where I do not let them decide.
Los payouts dejaron de despacharse. No hubo excepción, ni alerta, ni una sola línea roja en los logs: la flag que los gobernaba apuntaba a una key que ya no existía.
A customer had 97% of their balance locked and zero open operations. It was not a balance bug: it was a definition bug.
The provider said we owed them six figures. Our systems said we owed nothing. The debt existed nowhere at all.
How I use Python to read sensors, decide and move a robot — from the control loop to simulation.
Robotics teaches systems thinking, failing fast and respecting reality. Skills I use every day.
The value is not in the model, it is in the decision it enables. My practical approach to data science with Python.
From pixels to decisions: how computer vision lets a robot understand the world.
Status, risks and projection: the three signals I review so a project lands on time with no surprises.