The problem it addresses
AI is moving fast into fraud and payments operations. Models now write detection rules, suggest thresholds and take actions on live accounts. In most systems the model acts first and someone checks later: a reviewer samples decisions, a log records them, a second model grades the first.
In fraud, that order is costly. A payout that has left is gone. A good customer whose account was frozen in error pays for the mistake before anyone reviews it. And when an auditor or regulator asks why a decision was made, an account of what the model usually does is not enough.
Riskernel’s engine is designed so that checking comes first. The patent application covers that design across the full life of a fraud decision.
What the application covers
The application was filed with 215 claims. It covers the engine end to end, in five areas.
1. Admitting new detection logic
New rules and features enter the engine only after they pass checks. Logic that fails never reaches a live decision. Teams can adopt new detection ideas quickly and still know that each one was examined first.
2. Sharing what works across institutions
Fraud rings rarely attack one institution at a time. The engine lets what works at one institution help others, while each one keeps its own data. Knowledge travels, and customer records stay where they are.
3. Real-time decisions within fixed limits
Decisions are made in real time, within limits each institution sets in advance. The engine stays inside those limits on every decision.
4. Turning checked logic into fast models
Logic that has passed its checks becomes fast models that can run on live traffic. The speed comes with the assurance the checks provided.
5. A loop that keeps improving all of it
The engine learns from its own results and from new attacks. Each improvement goes through the same checks as everything else, so the system gets better while its controls stay in place.
What it means for a fintech or bank
- No AI action runs until it has been checked.
- Fraud patterns can travel between institutions while raw data stays with each one.
- Each decision stays within limits set in advance.
- Every decision can be replayed for an auditor or regulator.
- Detection keeps improving over time.
For a risk team, this brings the speed and reach of generative AI to a system whose behavior can be explained, bounded and defended.
Where it comes from
Riskernel combines algorithmic verification with generative AI, using the algogen framework Amir Shachar described in his 2024 paper (arXiv:2403.01426) and his book. In that framework, models only propose, and deterministic algorithms check every proposal before anything acts. The Algogens page explains the framework in more detail.
The application builds on a long record of invention. Amir Shachar holds 9 granted US patents and has 5 pending, including this one. His earlier work includes commercial federated learning at NICE Actimize, protected by granted patents. The full list is on the research and patents page.
What stays in the application
This page describes what the engine achieves. How it achieves it is set out in the application itself, which is not public.