Fraud caught in
11 milliseconds.
Before the cardholder's phone buzzes. Before the chargeback lands. Before the margin erodes. Sentinel's ML engine screens every transaction in real time — no rules, no thresholds, no lag.
Enter three numbers. See what fraud is actually costing you — and what Sentinel saves.
Three-layer ensemble.
One decision.
Sentinel doesn't run a single model. It runs three simultaneously — each specialized for a different fraud signal — and fuses their outputs in a calibrated meta-learner before your authorization switch fires.
Gradient-boosted trees score 140 velocity features — spend acceleration, merchant category drift, geographic displacement — in under 3ms.
128-token transaction history fed into a 6-head attention transformer. Catches account-takeover patterns invisible to rule engines.
Propagates fraud signals across the merchant-cardholder-device graph. Detects organized ring fraud before the second card is hit.
Plugs into your existing stack.
Not the other way around.
REST webhook, Kafka connector, or full SDK. Sentinel sits in-line before your authorization decision — no rerouting, no latency tax, no forklift migration.
import sentinel
client = sentinel.Client(api_key="sk_live_...")
# Score a transaction in real-time
result = client.score(
transaction_id="txn_9f3a2b",
amount=249.99,
currency="USD",
merchant_id="mcc_5411",
card_hash="sha256:a3f9...",
device_fingerprint="fp_7c2d...",
ip_address="104.21.8.42",
timestamp="2026-02-27T05:29:00Z"
)
print(result.fraud_score) # 0.003
print(result.decision) # "APPROVE"
print(result.latency_ms) # 9.2
print(result.shap_features) # [{feature: "velocity_1h", ...}]{
"transaction_id": "txn_9f3a2b",
"fraud_score": 0.003,
"decision": "APPROVE",
"latency_ms": 9.2,
"model_version": "sentinel-3.2.1",
"shap_features": [
{ "feature": "velocity_1h", "value": 1.0, "impact": -0.12 },
{ "feature": "merchant_trust", "value": 0.94, "impact": -0.08 }
],
"audit_token": "aud_8f2a9c4d..."
}
Meridian Bank
Meridian's fraud team was manually reviewing 4,200 alerts per day with a 12-person staff. Within 30 days of deploying Sentinel, alert volume dropped to 340 — all of them genuine positives. The team reallocated six analysts to proactive threat hunting.
NovaPay
NovaPay processes 600 transactions per second at peak. Legacy rule engine was blocking 2.1% of legitimate transactions — costing $28M in annual revenue leakage. Sentinel's precision model dropped false positives to 0.3% while catching 47% more fraud.
Every decision is
explainable.
Every audit trail
signed.
PSD3 requires Strong Customer Authentication and machine-readable dispute evidence from 2026. Sentinel ships SHAP attribution payloads on every score — the top five features, in plain language, ready for SAR filings.
47 Deployments.
Real Numbers.
No vendor math.
Sentinel runs a three-layer ensemble — a gradient-boosted tree for rapid feature scoring, a transformer-based sequence model for behavioral context, and a graph neural network for network-level anomaly propagation — all compiled to TensorRT and served on co-located GPU inference nodes within your cloud region. No cross-region hops. No cold starts.
Transactions scoring between 0.45 and 0.72 on our fraud probability surface enter a soft-block queue. Sentinel issues a step-up challenge (3DS2 or biometric) rather than a hard decline, preserving revenue while gathering additional signal. This reduces false-positive chargebacks by 67% compared to hard-threshold rule engines.
Median time-to-live is 14 minutes via REST webhook or our Kafka connector. Full SDK integration with your existing fraud stack (including Actimize, FICO Falcon, or Feedzai) averages 3 business days with a dedicated integration engineer.
Yes. Every decision includes a signed SHAP attribution payload — the top five features driving the score, in human-readable form — suitable for direct inclusion in SAR filings, regulatory audit trails, and customer dispute responses.
Your next chargeback spike
won't happen.
Median time-to-live: 14 minutes. Dedicated integration engineer included. No minimum contract.