Counterparty Credit Risk Systems
Software Engineer · JPMorgan Chase & Co. · 2018–2021
A credit-risk calculator spanning five countries, where a wrong number was a compliance problem, not a rounding error.
The Question
How do you know a risk number is actually right, not just plausible?
Context
I was responsible for the end-to-end functioning of a financial derivative product calculating counterparty credit risk across trades in the US, Korea, China, India, and South Africa, with stakeholders spread across 5 time zones.
The Constraint
Every calculation had to be right the first time. There's no 'ship it and iterate' when the output feeds a compliance report across five regulators in five time zones.
My Hypothesis
I assumed manual review by careful analysts would catch drift. It didn't scale. By the time someone spotted a mismatch by eye, it had usually already fed into a report.
What I Tried
Architected and launched the credit risk calculator itself across all 5 regions. Built an input-adjustment module for APAC specifically to cut down on manual intervention. Built an automated reconciliation tool whose only job was regression testing: catching when a calculation had drifted from what it should be.
The Decision
Invested in the reconciliation tool rather than relying on manual review to catch drift, treating "does this calculation still match" as something that should be checked automatically, not trusted by default.
The Build
An automated reconciliation tool whose only job was regression testing: flagging the moment a calculation stopped matching what it should, before a person had to notice.
The Evidence
100% compliance with client requirements across all 5 regions. The APAC input-adjustment module saved 100+ hours of manual work monthly. The reconciliation tool cut calculation-mismatch defects by 30%. As Head SPOC for the derivatives module, turnaround with international stakeholders improved by 24 hours. A cross-functional web portal I helped launch was used by 15 teams to monitor 6k+ runs a month for risk reporting.
What Changed
Manual intervention on APAC inputs dropped by 100+ hours a month, and turnaround with international stakeholders improved by 24 hours. Fewer discrepancies meant fewer rounds of back-and-forth.
What I Learned
This is where the instinct actually started: build the thing that checks whether the answer is right, not just plausible. Everything since (eval thresholds, blind scoring, SxS frameworks) has been some version of the same reconciliation tool.