CASE STUDY 2
Scaling a FinTech AI Team to Build Real-Time Fraud Detection Systems
AI TALENT STRATEGY CASE STUDY
A payments company processing several million transactions a day was bleeding money to a fraud pattern its rules engine couldn't catch fast enough. Fraud rings adapted faster than the quarterly rule updates could keep up. The fix needed a model scoring risk in under 100 milliseconds. Not a batch job. Real time, or nothing.
Chargebacks ran 40% above the board's tolerance. The rules engine handled known patterns well and new ones poorly, and by the time a rule shipped for a fresh scheme, the ring had already moved on. Latency mattered as much as accuracy. A model that hit 99% accuracy but tacked 400 milliseconds onto checkout was simply not shippable.
This is a narrower hiring problem than it first appears. It's not "find data scientists." It's "find people who've shipped a low-latency scoring model into a live payment path, and know exactly what breaks when you do."
We ran two tracks in parallel. One targeted full-time ML engineers with direct fraud, risk, or trust-and-safety experience at other payment or lending companies. The other used contract staffing to embed two specialists immediately: a feature engineering contractor who'd built real-time feature stores before, and an MLOps contractor who could stand up low-latency model serving without waiting on the permanent hires to onboard.
The technical screen leaned less on architecture, more on judgment. Every candidate got the same tradeoff: a new gradient-boosted model catches 12% more fraud but adds 60 milliseconds at checkout. What do you do? Strong candidates asked about false-positive tolerance, checkout abandonment data, and whether the latency budget could be recovered elsewhere in the request path. Weak candidates just answered the model question.
Roadmap:
The shadow-deployment window turned out to matter more than anyone expected going in. Running the model silently against live traffic for three weeks surfaced a class of edge cases, gift-card top-ups made from a new device on a shared account, that never showed up in offline testing against historical data. The permanent hires spent that window tuning specifically for those cases before the model was ever allowed to make a real decision.
Interviewers also began asking every finalist to sketch their own rollback plan before an offer went out: if the new model started rejecting legitimate transactions at scale, how would they know within the hour, and what would they turn off first. Candidates who had lived through a bad rollout answered in seconds. Candidates who hadn't tended to describe a plan that assumed nothing would ever go wrong.
Under 80ms at the 95th percentile
In the first quarter post-launch
3 permanent hires, 2 contractors, 1 contractor converted to full time
From kickoff
