Ensemble ML

Seven detectors.
One risk score.

Seven anomaly detectors feed a gradient-boosted meta-learner. Every score includes SHAP explanations and Bayesian uncertainty bounds for full regulatory transparency.

Request a demo →See the full platform
Ensemble snapshot

Calibrated scores. Quantified rationale.

Every score is calibrated, explainable, and bounded. SHAP attribution lands with each verdict so regulators see why, not just what.

SHAP
Explained scores
7
Anomaly detectors
29
Meta features
<200ms
Score SLO
Scoring Engine

Seven detectors. One ensemble. Full explainability.

Each transfer receives a calibrated risk probability from a 15-model ensemble. Regulators receive per-decision rationale with every score.

01

Anomaly detection

Statistical deviation, isolation forest, temporal patterns, velocity shifts, graph topology, dormant reactivation, and regime changes.

Detectors7 independent
02

Meta-learner ensemble

A gradient-boosted classifier stacks detector outputs with 29 behavioral features. Platt calibration produces true probability estimates.

Features29 dimensions
03

SHAP explainability

TreeExplainer produces per-feature attribution for every risk decision. Regulators receive quantified reasoning with each score.

ExplainedEvery decision
Pipeline

From transfer to risk score in four layers.

Detection, ensembling, and calibration happen in parallel. Every score arrives with deterministic explanation in under 200ms.

01·Extract

Feature extraction

39 behavioral features computed from transfer history. Velocity, timing, counterparty diversity, and graph metrics feed the scoring engine.
39 dimensions
02·Detect

Anomaly detection

Seven independent detectors run in parallel. Statistical deviation, isolation forest, temporal patterns, velocity shifts, graph topology, dormant reactivation, and regime changes.
7 parallel detectors
03·Ensemble

Meta-learner

A gradient-boosted classifier stacks detector outputs with 29 behavioral dimensions. The ensemble weighs each detector by historical accuracy.
29 ensemble features
04·Calibrate

Calibration

Platt scaling transforms raw scores into calibrated probabilities. Uncertainty bounds accompany each prediction.
Platt calibrated
Explain & deliver

Explain it. Deliver it.

Every score is paired with SHAP attribution and pushed to downstream agents within the same compliance pipeline.

05·Explain

Explainability

SHAP TreeExplainer generates per-feature attribution for every score. Regulators receive deterministic rationale for each decision.
SHAP explained
06·Deliver

Delivery

Calibrated scores write to the risk profile materialized view. Downstream agents consume scores within 200ms of transfer confirmation.
<200ms score delivery
See it run

See risk scoring produce a live verdict.

30 minutes. Real transfers scored live. Full SHAP breakdown for each decision.

Request a demo →See the full platform