I build AI for credit risk, fraud and compliance, then audit it the way a bank’s model risk team would. Three years inside regulated banking taught me that a confident wrong answer is the most expensive kind.
An agent that knows when it doesn’t know. It reviews loan agreements against the FCA Consumer Duty, cites its evidence and hands uncertain cases to a person.
Credit decisioning for thin-file BNPL customers that bureaus can’t see, on a synthetic portfolio I built: modelled, explained and governed. Built to pass a model risk review.
Fraud decisioning tuned to what fraud actually costs a bank, not a textbook metric, with an explanation for every decline.
Credit models tested on three things at once: performance, explainability and fairness.
I build the simplest version first and make the clever one beat it. My compliance agent had to outperform a fixed workflow before it got to choose its own tools.
A threshold is a business decision. I set them on the cost of a missed fraud or the capacity of a review team, never on a default 0.5 or F1.
Every major decision is logged with what I chose, what I rejected and why, including the audits of my own work.
years inside regulated banking: credit risk, compliance and financial data.
I came to AI from the inside of a bank. I assessed affordability for 2,000+ customers, audited 3,000+ accounts for compliance, and delivered 200% of my liability budget and 333% of my profit budget in 2023.
That’s why I build systems that can be evaluated, explained, monitored and trusted, not models that only perform in a notebook.
Today I build agentic AI, credit risk and fraud systems for regulated finance, with governance and deployment treated as part of the build.
Credit risk, fraud and explainable models built around the real cost of errors.
Retrieval-grounded agents that cite their evidence, check their own claims and escalate to people.
Pipelines that train, track, serve and monitor models, not notebooks that stop at a score.
Credit, KYC and compliance experience carried into every technical decision.
Tap a project above to highlight its skills · tap a letter on any skill to jump to that project.
Documenting the real compliance-agent build: what fails, what I measure, and what I have to fix before I can trust it.
Not one was classified as completely unsupported. That should feel like good news. It doesn’t.
Five manual tests suggested a clean document was flagged 1 in 5 times. A structured evaluation across 15 documents told the real story.
I built a check to catch made-up information, and almost trusted my own reasoning for why it would work.
Connecting drift monitoring to champion–challenger promotion, and the gap that appeared at the connection point.
Hiring for ML, AI or data science in financial services? I’d love to hear what you’re building.
korede@koredefolarin.com ↗