01
Project
Keep your evaluation in one project scope.
Create or select a project for data drafts, saved models and scoring history.
- Scope
- Named project
- Data
- Training drafts
- History
- Saved models and scoring runs
MAG OptiAIOperational AIFraud / Risk ScoringEarly Access
Score Risk Deliberately. Review the Evidence.
Prepare labeled data, train saved model candidates and score new records with a deliberate execution source. Review probabilities, evaluation evidence and feature contributions before deciding what needs attention.
Prepare labeled data. Evaluate saved models. Investigate scored records.
Prefer to explore first? Start free14-day trial · lower limits. Beta access requires approval.
Built for review teams
Labeled data and scored records
Cloud-based Pro Console
Guided evaluation
The risk-scoring challenge
A Score Needs Evidence and Context.
Labels, model evidence and review capacity shape how scoring can support your decisions.
Challenge 01 / 04
Reliable Scoring Starts With Reliable Labels
A model learns from the records and outcomes you provide.
Training needs usable features and a binary target. Check the class balance and data readiness, then ensure new scoring records match the saved model's feature schema.
Impact
Give training and scoring a consistent data foundation.
The challenge is more than producing a score.
It is keeping labeled data, model evidence and review context connected.
Why MAG OptiAI
A Clearer Way to Evaluate Models and Review Risk.
Keep training data, saved models and scoring evidence connected throughout the review.
A Consistent Data Foundation
Prepare labeled training records and validate scoring inputs against a saved model's feature schema.
A Deliberate Model Lifecycle
Train candidates, compare saved versions and choose which version to stage or execute.
Evidence for Review
Inspect probabilities, evaluation metrics and available explanations before taking action.
Your model and scoring context stay connected.
Saved projects hold training drafts, model versions and scoring history. Review the execution source and evidence behind the result you are investigating.
Process
How We Work
From labeled training data to saved models and records you can review.
MAG OptiAI guides your team through data preparation, saved-model evaluation and scoring with explicit provenance.
02
Data
Prepare the labeled training dataset.
Upload CSV or use samples, inspect the binary target, features and readiness.
- Input
- Features and 0 / 1 target
- Review
- Readiness and class counts
- Keep
- Named data draft
03
Models
Train, compare and stage deliberately.
Choose a model family, train saved candidates and inspect evaluation evidence.
- Train
- Saved model candidate
- Compare
- Versions and evidence
- Stage
- Explicit default selection
04
Score
Score records with the intended saved model.
Prepare feature-aligned rows and choose the staged model or a selected version.
- Input
- Feature-aligned scoring rows
- Execute
- Staged or selected version
- Optional
- Local feature contributions
05
Results
Review probabilities and their supporting evidence.
Scored rows · evaluation evidence · provenance · exports · grounded AI interpretation
- Review
- Probabilities and evidence
- History
- Open a saved scoring run
- Use
- Investigate and export
Core capabilities
What Fraud / Risk Scoring Can Do for Your Organization
A connected workflow for labeled data, saved models, feature-aligned scoring and investigation.
Prepare Labeled Training Data
Upload CSV or use samples and inspect features, class counts and readiness.
Choose a Model Family
Evaluate LightGBM, Logistic Regression, Random Forest or XGBoost candidates.
Keep Saved Model Versions
Train saved candidates and compare the evidence for different versions.
Scored records.
Review with evidence.
Score Feature-Aligned Records
Execute the staged model or a selected saved version with matching input features.
Inspect Evaluation Evidence
Review precision, recall, confusion-matrix counts, burden and available calibration evidence.
Review and Export Scored Rows
Inspect probabilities, saved scoring history and the model provenance behind results.
Stage a Model Deliberately
Choose a saved version as the default for future scoring.
Use Grounded Interpretation
Review available feature contributions and AI interpretation tied to the scoring result.
Prepare Labeled Training Data
Upload CSV or use samples and inspect features, class counts and readiness.
Choose a Model Family
Evaluate LightGBM, Logistic Regression, Random Forest or XGBoost candidates.
Keep Saved Model Versions
Train saved candidates and compare the evidence for different versions.
Stage a Model Deliberately
Choose a saved version as the default for future scoring.
Score Feature-Aligned Records
Execute the staged model or a selected saved version with matching input features.
Inspect Evaluation Evidence
Review precision, recall, confusion-matrix counts, burden and available calibration evidence.
Review and Export Scored Rows
Inspect probabilities, saved scoring history and the model provenance behind results.
Use Grounded Interpretation
Review available feature contributions and AI interpretation tied to the scoring result.
All in one place.
The tools to prepare, score and review in one connected workflow.
Fraud / Risk Scoring in action
See the Evidence Behind Scored Records.
Review model probabilities alongside their evaluation context and provenance. This illustration explains the concept; it is not a measured model result.
A probability deserves context before a decision.
Inspect scored rows, the saved model version and its evaluation evidence. Optional local feature contributions can help explain how the model produced a result.
Use scores to support investigation. Model probabilities and feature contributions do not establish fraud or prove causality.
Scored records and evaluation contextIllustrative
| Record | Probability | Evaluation label |
|---|---|---|
| Record A | 0.18 | Below threshold |
| Record B | 0.72 | Flagged |
| Record C | 0.39 | Below threshold |
| Record D | 0.86 | Flagged |
Who it’s for
Is Fraud / Risk Scoring a Good Fit for Your Team?
For teams evaluating a defined binary target and reviewing scored records in context.
Risk Review Teams
Evaluate labeled historical records and investigate scored cases against your operational review process.
Analysts Evaluating Models
Train saved candidates, compare evaluation evidence and choose a staged or selected execution version.
Teams Investigating Records
Inspect probabilities, available feature contributions and case context before deciding what needs attention.
Every dataset and review process is different. Evaluate model evidence and review burden against your own requirements.
The Fraud / Risk Scoring Evaluation
Real Problem. Real Evaluation. Real Impact.
Bring your risk-review question and evaluate the fit with your organization’s labeled records.
What You Get
Access for a Larger Evaluation
Apply through the existing beta process to evaluate beyond the smaller free-trial limits.
A Guided Data Setup
Frame the review question and prepare labeled training data and matching scoring records.
A Saved-Model Workflow
Train candidates, review evaluation evidence and choose the execution source deliberately.
A Direct Feedback Channel
Discuss data preparation, model comparison and scoring requirements that need a better fit.
Results You Can Investigate
Inspect probabilities and available interpretation to judge their usefulness for your review process.
What We Ask From You
A Defined Review Question
Describe the records, target outcome and decisions your team needs to support.
Usable Labeled Data
Bring representative features and a binary target. Anonymized records should preserve the relationships needed for evaluation.
Practical Feedback
Tell us where preparation, evaluation or interpretation needs improvement.
A Review of the Outcome
Assess model outputs against operational context and your organization's review requirements.
Real Problem
Real Evaluation
Better informed review.
A real-world evaluation with your organization.
Bring your risk-scoring problem and real or anonymized labeled data. We will review the fit and help you get started.
Fraud / Risk Beta
Frequently Asked Questions
What to know before evaluating your risk-scoring requirements.
What data do I need?
Training requires records with usable features and a binary 0 / 1 target. Review the workspace's CSV requirements and readiness checks. Scoring records must match the feature schema of the model you execute.
Which model families are available?
The current workspace supports LightGBM, Logistic Regression, Random Forest and XGBoost. Review each candidate's evaluation evidence against your data and review question.
Does training automatically stage a model?
Training creates saved model candidates. Use Stage Model to choose a default intentionally. Scoring can use that staged model or a selected saved version without changing the staged state.
How do I interpret evaluation thresholds?
The stored evaluation threshold provides a reference for result labels. Threshold sweeps and candidates are exploratory evidence. An evaluation reference, including 0.5 where used, is not an approved operational decision policy.
Can I review false-positive burden?
Evaluation evidence includes confusion-matrix counts and burden summaries alongside metrics such as precision and recall. Assess these against the cost of missed cases and your team's capacity for review.
Does a high probability prove fraud?
No. A score is model output that supports review. It does not establish fraud, prove causality or replace investigation and human judgment.
How does Early Access relate to the trial and beta?
Early Access describes product maturity. Start Free uses the existing 14-day trial with lower limits. Larger evaluations use the existing beta application and team approval; there is no separate access-request system.
What can explanations help with?
Request local feature contributions where supported and use grounded AI interpretation of successful scoring results. Feature contributions describe the model's output; they do not establish causes or make an operational decision.
How do I apply for a larger evaluation?
Use Apply for the Fraud / Risk Beta to reach the existing product-specific application. Our team reviews your requirements and confirms the fit for an evaluation.
Still have a question?
Talk to us about your risk-scoring evaluation.
Ready to get started?
Ready to Evaluate Fraud / Risk with Your Organization’s Data?
Tell us about your labeled records and risk-review requirements. We will review your application and confirm whether the current Fraud / Risk Beta is a good fit for your organization.
No payment requiredApplication reviewed by our team