AML Intelligence Case Study

How a regional bank can cut Suspicious Activity Report (SAR) turnaround from weeks to hours with our small language model that's fully auditable and examiner-ready.

Case Overview

Earlier risk detection, built to survive examination

Our models caught suspicious activity sooner, reduced false positive alerts, and preserved full lineage for regulatory review.

5.08M

IBM HI-Small transactions processed

66M

Classifier parameters

0.91

Validation ROC-AUC

500ms

Streaming scoring interval

Challenge

Rules-based screening missed the patterns that matter

Rules-based screening misses graph-shaped patterns like structuring, fan-out, and layering while analysts manually review 8-10 million monthly transactions across fragmented systems. High-risk alerts stall before disposition, and SAR submissions take 3-4 weeks from first trigger. Compliance leadership needs to surface suspicious activity earlier, without sacrificing the evidence quality regulators demand.

Measured Outcomes

Earlier detection, measured in the first production cycle

Counter-level metrics from the initial rollout period across compliance speed, review quality, and pattern coverage.

SAR turnaround time

4 hours

-95%

Before: 3-4 weeks

False-positive review load

12% lower

-12%

Before: Baseline

Pattern classes detected

4 topologies

+4 types

Before: Rules only

Validation quality trend (ROC-AUC)

Month 10.84
Month 20.87
Month 30.89
Month 40.91

Implementation Detail

Architecture used in production

The deployment keeps scoring, pattern extraction, and analyst decisions tightly coupled without sacrificing governance requirements.

Stage 01

Data Pipeline

Cleans IBM HI-Small source rows, normalizes timestamps and currency values, and yields reproducible fixtures for model training and replay.

Stage 02

Classifier Scoring

Scores transaction descriptors with class weighting and negative undersampling to keep the extreme class imbalance trainable.

Stage 03

Pattern Engine

Detects structuring, fan-out, fan-in, and cycle motifs using graph windows and emits explainable evidence bundles for analysts.

Stage 04

Review & Retrain

Persists analyst dispositions and promotes nightly retrained models into the registry with threshold and metric lineage attached.

4 weeks rollout timeline

Week 1

Data ingest and baseline benchmark

Connected transaction exports, profiled class imbalance, and validated baseline rule precision/recall envelope.

Week 2

Model fine-tuning and thresholding

Fine-tuned the scoring model and calibrated alert thresholds against validation cohorts and review capacity.

Week 3

Graph detector integration

Integrated topology detection and analyst pattern context into queue-level triage workflows.

Week 4

Production rollout and governance handoff

Activated governed model version, enabled audit-ready dossier export, and scheduled retraining cadence.

Need AML coverage for your data?

Use this same deployment pattern to modernize SAR operations, accelerate analyst throughput, and improve model governance confidence.