Continuous retraining
Every confirmed dispute outcome — won or lost — retrains the pre-transaction model. No manual retraining schedule. The feedback loop from chargeback outcomes to fraud detection is automatic and continuous.
SENTR · PLATFORM · AI RISK ENGINE
EU AI Act Article 13 compliant — by architecture, not by retrofit
A fraud scoring engine that explains every decision — not just produces a score
Gradient-boosted trees, sequence models, and graph neural networks — ensemble architecture trained on EU payment data, explainable at the feature level for every automated decision.
How the scoring layer is built
Three model types. One ensemble. Every transaction scored in under 100ms.
| Model Type | Purpose | EU AI Act Status |
|---|---|---|
| Gradient-Boosted Trees | Pattern detection — identifies known fraud typologies across transaction features | Explainable at feature level |
| Sequence Models | Behavioural timeline analysis — surfaces velocity abuse, account takeover patterns | Sequence attribution output |
| Graph Neural Networks | Entity relationship analysis — maps synthetic identity clusters, multi-account rings | Graph path explanation |
What the graph layer catches
Gradient-boosted models score individual transactions — velocity, device mismatch, IP flags, payment history. Every account in a fraud ring looks clean at that level. SENTR's graph neural network layer operates on the cross-account relationship graph: shared device fingerprints, overlapping IP ranges, and coordinated transaction timing reveal fraud rings and synthetic identity clusters that the individual transaction view was never built to see.
All three run in ensemble on every transaction. Score returned in under 100ms. Every score includes the top contributing signals from each model layer — returned in the API response, not generated post-hoc.
Feature-level explainability — not post-hoc rationalisation
Every score returned by the API includes the top contributing signals that produced it. This is not a SHAP value layer bolted onto a black-box model — the models are built with explainability as an architectural constraint, not an add-on.
EU AI Act Article 13 requires transparency for high-risk automated decision systems. SENTR satisfies this at the decision level, not the aggregate model level — every individual transaction decision has a structured, human-readable rationale available on demand.
{
"score": 0.87,
"confidence": "high",
"top_signals": {
"velocity_7d_cross_merchant": 0.34,
"device_fingerprint_mismatch": 0.28,
"ip_country_card_country_delta": 0.19
},
"decision_ref": "txn_ae4b91c2",
"eu_ai_act_exportable": true
}Model retraining and drift management
Model versioning
Every model version is logged with training date, data window, and performance delta vs. prior version. Version history is available on request for audit submissions and regulatory reviews.
Drift detection
Score distribution is monitored continuously against baseline. Drift alerts trigger before model performance degrades — not after chargebacks signal it. No silent model decay.
August 2026 — EU AI Act high-risk decision documentation requirement
SENTR's explainability architecture produces the Article 13 compliance documentation as a byproduct of normal operation — there is no separate compliance workflow. The audit log for every automated decision is exportable on demand, structured for regulatory submission, and maps to the SENTR decision reference ID returned in the API response.
- ✓ Feature-level explanation per decision
- ✓ Exportable audit log — structured for regulatory submission
- ✓ Decision reference ID traceable to full case record
Talk to the engineering team — not a sales deck
Architecture Sessions are technical calls with SENTR engineers. Bring your stack, your integration questions, and your EU AI Act documentation requirements. We'll walk through the model architecture, the API spec, and the integration timeline. See the full 8-day integration timeline →
Book an Architecture Session →