Use casesFraud and risk scoring
Catch the transaction that does not belong, and know how often you are right.
Fraud and risk are the first thing financial services asks of AI and the last place a wrong number is forgiven. We build scoring models that run inside your warehouse, rank every transaction, claim or application by how much it looks like the ones that went bad, and are graded the honest way: on months they never saw, against the simplest rule you already use.
What gets in the way
A chargeback lands six weeks after the sale. An application looks fine until the third payment is missed. By the time a case is confirmed, the pattern that would have caught it is buried in a year of records.
Rules catch yesterday’s fraud and flag a thousand honest customers to do it. A model vendor shows a score of 0.95, and nobody asks how the test was built.
The real question is not whether a model can find fraud. It is whether you can trust the number it reports, and whether your own team can keep it honest after the vendor leaves.
What we build
The model scores. Your rules decide. Both live where your data already is.
- A point-in-time feature pipeline rebuilt nightly from your source tables as they stood that night, so no signal that leaks the outcome survives into training.
- A score for every transaction, claim or application, written once to one table with the signals that drove it, so any decision can be audited against exactly what the model knew.
- The threshold lives outside the model, in your rules engine or your case system. Risk can tighten it on a Monday without a deploy.
- A separate anomaly layer watches for what the model has never seen: a new merchant pattern, a volume collapse, a surprising new share. It routes that to a person rather than scoring it with false confidence.
- Drift monitoring that checks scores against confirmed outcomes as they arrive, and a retraining notebook your own data team runs.
How we know it works
We grade a model the way it will be used. It learns from the past and is tested on the months after.
- An earlier vendor reported a 0.95 to Winland Foods by shuffling every month into one pile. Graded honestly, on months after the training cut-off, the score was 0.81. Yours will be reported the honest way, with the simple rule it has to beat printed next to it.
- Before launch, every historic case is rescored through the production pipeline and compared with development. Winland’s 7,709 matched with a correlation of 0.998.
- The anomaly layer is Verdict, our accelerator, which ranks the right cause first 86% of the time on a public benchmark and says so when it finds none.
- Everything a model validator asks for comes out of the build: time-split validation, the baseline comparison, the feature list with leakage checks, and a score written once per decision.
Built from this work
We have not shipped a fraud model under a bank’s or an insurer’s name yet. The scoring spine, the honest grading and the anomaly layer are all live in the work below.
28% of claims clear automatically at the recommended bar
92% of cleared claims were real shortages, on months the model never saw
$174K in deductions cleared through the model at the first month-end close
Read the story86% right cause ranked first on a public benchmark, against 14% by chance
8.7% right cause first on a simulated chemical plant, our open problem
10,300 labeled anomaly windows behind the measurements
Read the story
Where it fits
Insurance
Insurance, from the question a customer asks to the systems that answer it.
Financial services
AI for financial services, built to pass an audit.
Questions
How do you handle the flood of false positives?
By reporting precision at every threshold on held-out months, so you choose the trade between caught fraud and flagged customers with the numbers in front of you. The threshold is yours to move.
Does it need labeled fraud cases?
It needs confirmed outcomes: chargebacks, write-offs, confirmed cases. Where labels are thin, the anomaly layer carries more of the load and the model’s reported confidence drops to match. We do not pad it.
Where does the data live?
In your warehouse. The pipeline, the model and the scores table all run there, under your access controls. Nothing leaves.
What about model risk management?
The build produces what a validator asks for: time-split validation, the baseline comparison, the feature list with leakage checks, and a score written once per decision with the signals behind it.
Bring us the project that matters most.
Tell us what you’re building and what’s in the way. You’ll hear back within 24 hours.