Essay

Why fraud models should only propose. By Amir Shachar.

Read the essay
Skip to content
Riskernel
Patent pendingFraud engine for fintechs and banks

Generative intelligence. Algorithmic certainty.Every AI decision is checked before it acts.

Riskernelâ„¢ scores and explains payment and payout risk with a patent-pending engine built on the Algogens architecture framework, in which machine learning, generative AI and deterministic algorithms each have a fixed job, and nothing acts until the algorithms have checked it. Fraud patterns travel between clients through federated learning while no client’s raw data ever leaves their boundary.

Runs beside your current system. Nothing to rip out. Scores at a 47 ms median (Riskernel-reported; single calls vary).

IncomingWireAmount2,300 USDPayeeNew

Learnpatterns learned locally, hardened red vs blue

Patterns travel between institutions. Raw data stays put.

Proposemodels score it; borderline payments get an AI proposal

risk0.91deviceremote-access appsequencecall in progress

Push-payment scam likely. Freeze the account.

Checkno model judges

thresholds learned offline, then fixed

  • Schema integritypassed
  • Telemetry groundingpassed
  • Policy consistencypassed
  • Blast radiusfreeze also stops 2 payroll payeesfailed
  • AI revises, checks rerunAI revised: hold this payment only
Actonly actions on your listActed in 48 msVerify callerDeterministic validator: 4 of 4 passed after 1 repair

Machine learning and generative AIDeterministic algorithms

Illustrative trace with sample data.

Exhibit A · The record

Pioneered commercial Federated Learning for Tier-1 fraud prevention.

Amir Shachar led the R&D and patented commercial deployment of Federated Learning across global Tier-1 financial institutions at NICE Actimize. Riskernel builds on that foundation and the AlgoGens framework.

  • 2019

    CEO Excellence Award at NICE Actimize, for patented commercial deployment of federated learning.

  • 9

    Granted US patents (plus 6 applications): machine learning, AI, fraud detection.

  • Tier-1

    Former Head of Fraud Data Science at NICE Actimize, deploying models for Tier-1 institutions.

  • Research

    Author of research monographs on Hybrid AI Architectures and Rare-Event Stream Modeling.

Algogens

Generative intelligence with algorithmic certainty. Models only propose. A deterministic validator checks every proposal before any action runs. A failed proposal is repaired within bounds or rejected.

Federated learning

Fraud patterns travel between clients, and no client’s data ever reaches another.

Cost

Proof first. Budget later.

Fraud defense should not need an enterprise budget, and it should never need one before you have seen a result. Riskernel shows the evidence first, on your own data, free of charge.

Evidence on your data.

A free offline evaluation of one flow on your historical data. The written report arrives within 30 days after the data does.

One endpoint to integrate.

Scoring is a single API call that returns a risk score and the reasons behind it. It runs beside your current system, with nothing to rip out.

Research you do not have to fund.

Riskernel is built by a named inventor on 15 patent filings, 9 of them granted US patents. A team your size gets that work without hiring for it.

If the report does not earn a next step, it ends there and your data is deleted.

Launch Fraud Review

What are you launching next?

A new payout rail, a card program, a lending product, a new country. Tell us in two lines. You get a private attack tree back: how the new flow gets abused, what it could cost you, and which three controls matter most. Signed by Amir Shachar. No call needed.

What you send stays private. One signed reply from Amir, no mailing list.

Runs beside your current system. Nothing to rip out.

  1. 01 Launch Fraud Review Free, in writing

    Private attack tree of the new flow

    Exit: Keep it. No obligation.

  2. 02 Shadow test 14 days, free

    Scores and reasons on your own traffic, next to your current results

    Exit: Keep the report. No obligation.

  3. 03 Architecture sprint 10 days

    Written review of your fraud surface, scoped to one number we name with you first

    Exit: Plan in hand.

  4. 04 Pilot 30 days

    Engine live in parallel on one flow, measured against that number

    Exit: Go or no-go on evidence.

  5. 05 Production Ongoing

    Real-time scoring API sized to your traffic

    Exit: Quoted on request.

Exhibit B · Algogens

Every AI in this engine is checked before it acts.

Other vendors let a model decide and ask a person to look later. Riskernel works the other way: the model proposes, a deterministic validator checks, and nothing executes until it passes.

Algogens vs agent-in-the-loop
Agent in the loop (typical)Riskernel Algogens
Who proposesThe model, which also actsThe model, which can only propose
Who verifiesA person after the fact, or another modelA deterministic validator, before anything runs
What a regulator can replayThe outcomeThe full decision path, step by step

Research behind the engine

9 granted US patents, one engine

Fifteen patent filings in total (9 granted US patents and 6 applications) across machine learning, privacy-preserving AI, and fraud detection. The Riskernel engine itself is patent pending, and three of the five earlier lines of work show up in how it is built.

The Riskernel engine: Algogen decision systems

In the engine This is the engine itself. Riskernel is built on Algogens, and one patent application, filed with the USPTO in November 2026, covers its whole decision lifecycle.

  • The full decision lifecycle How fraud logic is proposed, checked, shared and run One application covers the engine end to end, from the moment a new piece of detection logic is proposed to the moment it makes a live decision, and then back around again as the system learns. It was filed with 215 claims, and Amir is the sole inventor. Why it matters for fraud teams Every action is checked before it runs. Each client’s data stays with that client. Every decision can be replayed later, so an auditor can see what was decided and why. Read about this patent
    What the application covers
    1. 1Checked before it runs New logic is admitted only after it passes checks.
    2. 2Shared, not pooled What works helps every institution; data stays put.
    3. 3Real-time, within limits Live decisions stay inside fixed limits.
    4. 4Fast models from checked logic Checked logic is turned into fast models.
    5. 5A loop that keeps improving The system improves, under the same checks.
    US provisional 64/197,364 patent pending

Responsible AI

  • Fair and unbiased models Checking fraud models for bias before they go live Trains a fraud model on data balanced across attributes such as gender, age and location, so it is not skewed toward some groups. A bias tool then scores the model, and it is deployed only if the measured bias is low enough. Status A published US patent application (2023), filed at NICE Actimize.
    How it works
    1. 1Balance The training sample is balanced across non-sensitive attributes.
    2. 2Train The fraud model learns from that sample.
    3. 3Measure A bias tool scores the model’s risk predictions.
    4. 4Gate Below the bias threshold the model deploys; otherwise the sample is rebuilt and the model retrained.
    US 2023/0306429 published application

Learning across clients: federated, transfer and meta-learning

In the engine Riskernel is built on Algogens and federated learning: models only propose, deterministic validators decide what runs, and fraud patterns travel across institutions without data exposure.

  • Federated + transfer learning Anomaly detection that learns across institutions Shares learned signal between entities while each keeps its own transaction data. How it works Small models trained on a separate dataset produce risk scores for the transaction data. Those scores are added to the data before the main anomaly model is trained on it. US 12,124,933 granted
  • Recursive federated learning Automatic fraud models for each new client Stacks each client’s model on what the consortium already learned. How it works A model is trained for every client from its own data. A second model for each client is then trained on the first model’s scores, with stacking applied again. Every model outputs a score, so each stage feeds the next. US 2023/0237494 pending
  • Transfer learning + clustering Fraud detection when a client has almost no fraud yet Clusters the data, picks features by their explanations, and transfers signal from elsewhere. Why it matters New clients often have too few fraud cases to train on. This approach compares a client with other clients and ranks which features to keep. US 2023/0267468 pending
  • Meta-learning A training engine that learns how to train Trains a model to predict the best learning rate at each training step. How it works It first simulates training runs and records the learning rates they use on the way to a minimum. Those records teach the model, which then predicts the rate for each step in place of gradient descent. US 12,481,890 granted

Fraud detection at scale

In the engine The same discipline runs through Algogens: machine-made rules and models are tested before they act, and a deterministic validator checks every proposal before any action runs.

  • Extreme class imbalance Supervised and unsupervised models, combined An ensemble built for client datasets where fraud is extremely rare. How it works It measures how rare fraud is in other clients’ datasets and sends each one to a supervised or an unsupervised model, depending on a threshold. The models are then combined into one fraud score per transaction. US 12,541,721 granted
  • Real-time scoring at scale Less memory, faster fraud decisions Keeps compact quantile summaries per customer instead of full histories. Why it matters Scoring needs a profile of each customer’s normal values. Storing a few marker values instead of every value cuts memory and speeds up processing. US 12,118,558 granted
  • ML-generated rules Decision rules written by machine learning Generates and validates alert rules for a fraud policy engine. How it works It generates candidate rules with machine learning, tests them, and drops the weak ones. Rules are picked using alert metrics, so the final ruleset suits the task. The output is a decision ruleset for a machine learning job. US 2024/0394713 pending
  • Model testing in the cloud Testing each client’s model before it goes live Validates a tenant’s classification model inside a shared cloud service. How it works The model is scored on the datasets of similar financial institutions on the same service. It goes live only if the average score clears a set threshold. Scores from each institution’s data are stored first, then averaged. US 12,626,151 granted

Deep and self-supervised learning for security

In the engine The same question shows up in the validator: how far could an action reach? Blast-radius solvability is one of its four checks.

  • Self-supervised learning Catching attacks while they still look harmless Learns normal cloud activity without labels and flags early signs of malicious behavior. How it works Unsupervised learning first marks which activity sequences may be malicious. A second model, trained on that, predicts whether a new sequence will lead to a malicious action. US 12,476,990 granted
  • Security posture management Cloud posture checks fed by live threat data Uses language models to summarize each company’s cloud security events and to find unusual patterns across companies. How it works Those patterns are compared with another company’s current cloud setup, and a better security posture is recommended for it. US 12,634,298 granted
  • Attack-path threat assessment Scoring threats by the path an attacker is likely to take Rates risk from the routes attackers are expected to follow toward valuable assets. How it works It estimates how likely an attacker at one cloud asset is to reach another along each path, then combines the likelihoods into a risk score. The European filing of the same invention is EP 4744195. US 12,425,439 granted
  • Attack-path threat assessment Scoring threats by the path an attacker is likely to take Rates risk from the routes attackers are expected to follow toward valuable assets. How it works It estimates how likely an attacker at one cloud asset is to reach another along each path, then combines the likelihoods into a risk score. This is the European filing of the same invention as US 12,425,439. EP 4744195 pending

Behavioral statistics

  • Signal divergence Flagging items whose clicks and real engagement disagree Compares click-through rate with downstream behavior such as time spent and conversions. When a title draws clicks but the content behind it does not hold readers, the signals disagree and the title is marked as a candidate for replacement. What else the claims cover Suggested wording is limited to titles in the same category as the one being replaced. The revised page is then served to users, so the change is tested on real readers. The independent claims cover a method, a system and a storage medium.
    How it works
    1. 1Measure Engagement is tracked for each page and, separately, its title.
    2. 2Compare Page and title engagement are checked against separate thresholds.
    3. 3Suggest Wording comes from high-engagement titles in the same category.
    4. 4Confirm New engagement data checks that the change helped.
    US 10,607,253 granted
1 / 6

Exhibit C · Public teardowns

How launches get abused, mapped in the open.

We take public launches and incidents and draw the attack tree: how the flow gets abused, what we can confirm from public sources, and what we cannot yet.

GOAL Cash out accounts the attacker does not own
  1. AGet valid logins
    1. A1Credential stuffing with pairs leaked elsewhere
    2. A2Buy ready-made hijacked accounts from a “shop”
  2. BGet through the front door at volume
    1. B1Automated login tooling and large wordlists
    2. B2Test: does a login from a new device need extra proof?Hypothesis
  3. CMove the money out
    1. C1Add a new payment method, make a five-dollar verification deposit, withdraw everything
    2. C2Test: withdrawal to a method already on the accountHypothesis
  4. DMonetise access
    1. D1Sell accounts in bulk to a reseller
    2. D2Sell single accounts through their own shop
From teardown 001, DraftKings 2022: how stolen logins became withdrawals. Built from public reporting (SecurityWeek, BleepingComputer, Help Net Security); court details are second-hand through those outlets. High, Medium and Low are our qualitative judgement of likely loss per leaf, not dollar estimates.

Two of the teardown’s seven findings

Confirmed

F-03 Confirmed

The cash-out route was a new payment method plus a five-dollar verification deposit.

Per the complaint as reported by BleepingComputer, buyers were told to add a new payment method, deposit five dollars to verify it, then withdraw all existing funds to a separate account. Help Net Security says that “in some cases” the group added payment methods under their control.

Likely loss
High
First control
Treat “new payout destination” as a high-risk event. Hold or cap the first withdrawal to any newly added destination.

Not yet confirmed

F-06 Not yet confirmed Hypothesis

Test: does a first withdrawal to a newly added method wait for a hold or a step-up check?

No source we read says what checks ran between “payment method added” and “withdrawal paid.” We would test whether a first withdrawal to a new destination is delayed, capped or challenged, and whether the verification deposit itself is a signal.

Evidence
No public source states which checks ran. Sources read: SecurityWeek, 20 December 2022 · SecurityWeek, 24 June 2026

Read the teardowns

Start with the attack tree. Stop at any rung.

Every rung before production ends with something you keep, and each one names its exit up front. Runs beside your current system. Nothing to rip out.