Financial-crime operations you can prove, inside your own perimeter.

Computed detection: real trained models, not seeded demo data

At the core of Guardian are two real, trained scikit-learn models behind a stable scoring interface. A gradient-boosted classifier scores transaction fraud (structuring bands, high-risk corridors, dispersion, above-CTR cash, velocity). A calibrated classifier scores entity / KYC risk from aggregate behavior and sanctions exposure. Scores are computed from your actual Transaction and Entity rows. The detection pass aggregates each entity's history, scores every transaction, flags the ones over threshold, recomputes the entity's risk score, and raises an alert from the strongest signal. Nothing here is hardcoded in seed data anymore. The alert you see was computed from the data underneath it.

On a held-out synthetic test split, the transaction-fraud model reaches ROC-AUC approximately 0.90 and the entity-risk model ROC-AUC approximately 0.95, with full precision/recall/F1 recorded in a versioned model card. Those are measured numbers from a real training and evaluation pipeline, not marketing targets.

Real sanctions and PEP screening

Screening runs against a committed subset of the public OFAC SDN list. A real data file that is parsed, normalized (case, diacritics, legal-suffix and honorific stripping), and matched with token-overlap plus fuzzy sequence similarity to produce a confidence score and a clear match status (no match, potential, confirmed). A DOB corroboration step nudges confidence where it lines up. A pluggable live-provider adapter (ComplyAdvantage, Refinitiv World-Check, Dow Jones shape) sits behind the same interface and fails closed. It raises rather than fabricate a result when no vendor credentials are configured. You screen against a real list today, and wiring a live feed later is a config change, not a code change.

UBO and money-flow graphs from real rows

Guardian's entity graph is traversed from your database, not drawn from fixtures. Ownership mode walks the KYB application's beneficial owners and directors, screens each node, and flags ownership above the 25% disclosure threshold and any sanctions/PEP hits. Transaction mode aggregates an entity's real counterparties by volume, screens them, and flags flows to sanctioned parties or high-risk (FATF) jurisdictions. When there is no data for an entity, the graph is honestly small. Never an invented scenario.

One workflow: alert to case to signed SAR

Detection raises an alert with a model explanation and the transactions that drove it. Analysts triage it, open and assign a case, and compose a SAR. Guardian auto-maps FinCEN fields from the narrative, then renders a BSA E-Filing v2.4 SAR XML envelope. On sign-off, an authorized signer applies a SHA-256 signature, the XML is generated and stored, and the filing is downloadable for submission. One queue, one case, one filing. No swivel chair.

On-prem isolation as the default

Guardian is single-tenant and on-prem-first. It deploys inside the institution's own perimeter, supports air-gapped installs with internal-CA certificates, and keeps admin surfaces reachable only over your network. Your data never leaves your walls. And the platform is cloud-portable later, per service, without a rewrite.

Audit and access control, by construction

Every state-changing request lands in an append-only audit log written in the same transaction as the change it records. Actor, role, resource, outcome, and a structured diff. Access is governed by role-based controls for the roles compliance teams actually run. Fraud investigator, BSA officer, IT auditor. With MFA and OIDC authentication. The evidence an examiner asks for is a byproduct of normal operation, not a separate project.