Turn signals into decisions.
Explore transaction risk with transparent scoring, a live review queue and synthetic ground-truth metrics. Built for learning, not real-world payment decisions.
Risk distribution
Active synthetic transactions grouped by rule-based score
Primary risk signals
Number of active records with each contributing signal
Priority investigations
Highest-scoring active records needing analyst attention
Review with context.
Every score has an auditable explanation. Change case statuses without affecting the model's synthetic ground-truth label.
Explore detection trade-offs.
Change alert cutoffs and simulated attack conditions. Precision and recall are calculated only against fictional demonstration labels.
Scenario presets
Apply synthetic changes to the same stored transactions
Risk rule controls
All modifications are visible and reversible
Scenario only: Amount, velocity and failure adjustments change computed scores, but do not modify stored transaction inputs or fictional truth labels.
Synthetic confusion matrix
Alert decision versus simulated fraud ground truth
If the denominator is zero, precision or recall is shown as N/A rather than 100%.
Threshold response
Precision, recall and alert volume across candidate cutoffs
Manage the sample ledger.
Add, edit, archive or restore transactions. Everything stays in this browser unless you explicitly export a file.
Active records
Model inputs, labels and review statuses
Archived transactions
Archived records are excluded from all current calculations
Workspace controls
Import validated data or reload the illustrative demo
Only FraudLens version 1 backups are accepted (maximum 300 records and 1 MB). Import replaces this local workspace only after confirmation. Export a backup first if you wish to keep your changes.
Every point is explainable.
This demonstration deliberately uses fixed, readable rules. It does not claim to be trained on real fraud data or to detect actual financial crime.
Risk features and points
How metrics work
A transaction is considered an alert when its computed risk score is at or above the adjustable cutoff.
- Precision = TP / (TP + FP), when alerts exist.
- Recall = TP / (TP + FN), when labeled fraud exists.
- False positives are alerts on simulated legitimate cases.
- True labels are independent, synthetic demonstration annotations.
Thresholds and signals do not learn from analyst actions or archived records.
Limits and safeguards
- All transaction identities, merchants and fraud labels are fictional.
- Rule-based scoring is not a calibrated likelihood or probability.
- There is no API, model inference, external service, or payment processing.
- Local browser storage is not encrypted or suitable for sensitive personal data.
- Scenario results are illustrative, not out-of-sample validation.
Portfolio implementation
Built for Hugging Face Static Spaces with one self-contained HTML file, responsive layout, SVG graphs, offline JavaScript scoring, CSV export, validated JSON import and optional dark mode.
Portfolio owner: Hugdrigo
Project number: #90
License: MIT
FraudLens © 2026 · HUGDRIGO / PROJECT 090 · Demonstration software only. Never use sample scores to make real financial decisions.