GPU ML Pipeline on AWS with Terraform
I built this Terraform-managed AWS ML pipeline to run GPU and CPU jobs on demand without paying for idle compute. In this post, I walk through the architecture, the key design choices, and the operational pitfalls to watch for before using this pattern in production.
Approved Is Not Paid
Starting from the simplest possible model, two parties and one ledger, this post derives the full architecture of cross-bank payments through a series of questions. It maps the 7-layer execution pipeline, identifies the cross-cutting concerns that prevent it from breaking, shows who controls each layer and why the settlement layer is intentionally closed to private actors, and traces a single card payment end to end. The central insight is that authorisation and settlement are different events, that regulation is load-bearing rather than incidental, and that the system runs on delayed consistency held together by institutional trust.