BNP Paribas

Summary

I’m a software engineer in BNP Paribas’s risk division in London. I work in C# on the engines that price and risk-manage trades on CPU and GPU, and I’m the team’s only C# engineer.

The role

The engines run large simulation-based calculations across several asset classes. It’s a big codebase, around 70 projects, and some of the C# is translated to run on the GPU.

I joined in July 2025 having never written C#. Today I own implementation and delivery on the C# side, from planning through review.

The problems

• Market data. Loaders written in Java had to be ported to C#, pulling data from several sources into one place and checking it on the way.
• A stale dependency. The live pricing pipeline relied on a data file that was no longer reliably available in production.
• Quant code. Pricing code written by quantitative developers had to be integrated into the engine.
• Two execution paths. The same tests had to run on both CPU and GPU.

Decisions

• Ship the risky change behind a flag, switched off. Removing the dependency touched live pricing, so I put the new path behind a feature flag that defaults to off, added caching, and wrote an end-to-end test nobody had asked for. The architect’s view was that the thing to protect was the blast radius, not only correctness. I’d have called that a nice-to-have before; now it’s how I start.
• Validate the data before trusting it. The commodity loader was the harder one: ten file types, correlation groups, unit conversions, and records linked to each other. I checked the links, so an orphaned record gets caught on load rather than showing up as a strange number in a simulation.
• Give AI work phases. I built a workflow where each stage has its own scoped agent: grill the problem, plan, build test-first, self-review, then a health review. Each agent only sees what that stage needs, which keeps it from skipping ahead.

What changed

I used that workflow to integrate a pricer into the calculation engine in a couple of days. For an experienced integration engineer it takes about two weeks, and I was new to it.

I went from no C# to shipping production code independently in six months.

The Early Careers digitisation competition is the other half of this. I led a team of four and co-built a dashboard from our users’ requirements, which cut a 30–40 minute data load to under 5. We won, out of five groups, and presented it to the Head of Risk Systems.

What I took with me

I was nominated as an AI Champion in the risk division, and the workflow is now used by several teams across the department. I also won a Take Ownership award.

The habit underneath all of it is simple: write down what you’re doing before you build it, and ship it so it’s safe to undo.