← All posts
Sep 18, 20262 min read

Ten years of risk systems: lessons from banking software

careerjavafintechleadership

In 2007 I walked into a small office in São Paulo as an intern and became employee #1 of a company that would spend the next decade building risk management software for Brazilian banks. When I left in 2016, the platform had five systems, a team of developers I had hired and led, and a credit risk engine that processed millions of records. Here’s what that decade taught me.

Employee #1

The first version of the platform was written in PHP by a very small team learning as we went. It worked well enough to win the first customers — which is a form of success that people who start with perfect architecture never experience. You can’t learn what a product should be until real users are angry about what it isn’t.

By 2008 we had enough scar tissue to rebuild it in Java. The second version corrected the mistakes of the first: proper layers, a real data model, and a build that didn’t depend on one person’s laptop.

The Java rebuild

The rebuild wasn’t a big-bang rewrite of everything. We took the five systems one at a time, kept the contracts that customers depended on, and replaced the internals. Some principles from that era are still with me:

  • Optimize where the data is, not where the code is — half of our wins were SQL
  • A versioned domain model pays for itself the first time a bank changes a regulation
  • Continuous integration isn’t a tool purchase, it’s a habit you have to defend

Leading the team

From 2014 I coordinated the development teams: agile ceremonies, architecture reviews, hiring. The hard part was never the code — it was helping developers with ten years’ seniority on me unlearn habits the platform had outgrown.

Legacy systems aren’t bad code. They’re the sum of every decision a team made under constraints you never saw.

What I’d redo

  • I’d push for automated tests of the risk calculations years earlier; spreadsheets as the test oracle don’t scale past one analyst
  • I’d treat the deployment process as a product feature instead of an afterthought
  • I’d challenge the custom framework (VRaptor) decision earlier — it worked, but it made every hire slower

Why legacy is a people problem

Everyone who joined the company in its first year had taught themselves parts of the domain. Everyone who joined later had to learn it from us. The code survived because the knowledge kept moving between people, not because it was well documented. The platform’s real architecture was a conversation.

I’ve carried that with me everywhere since: documentation is a nice-to-have, but teams that talk to each other about the system they own are the only ones whose legacy code keeps working.