I fixed broken payment systems live at 2am.
Now I build AI that survives production.

For ten years my job was to know how money actually moves. SWIFT, Lynx, ISO 20022, ACH, cross-border settlement, across HSBC and Société Générale. Not the marketing version of payments. The 2am version: the batch that half-processed, the one malformed record that failed an entire file, the reconciliation that broke three systems downstream because someone truncated a remittance field a decade earlier.

That work taught me a discipline that has little to do with technology being clever and everything to do with respecting how it fails. Scope before you fix. Find the poison pill. Race the cutoffs. Tell people early, because clients forgive a delay and never forgive silence. Write the root-cause analysis, because it is the only thing standing between you and the same incident twice.

What changed is that I stopped only writing about this and started building. I now design and ship production AI systems against real problems: multi-tool agent integrations, scheduled autonomous workflows, a full-stack app on a live database that my own household depends on. Not demos. Things that run real work end to end, break in real ways, and get fixed.

The surprise was how familiar it felt. Running a fleet of AI agents is the same job as running payment operations. Schedule the jobs so they do not step on each other. Build idempotency so nothing posts twice. Keep a hard respect for the limits of the system, except now the budget is a context window instead of a settlement cutoff. Ten years of operations turned out to be ten years of training for this.

I write here a few times a week about the infrastructure underneath agentic payments: how autonomous systems will authenticate, transact, and settle, from someone who operated the rails and now builds on top of them. If you are putting AI into production against a real operational problem, join the waitlist or join the weekly newsletter.