Enterprise AML Compliance
Without the Enterprise Price Tag
SwiftIQ Guard: a privacy-first, on-premise SWIFT message screening platform using a 4-tier AI decision chain - deterministic rules to on-premise LLM - at 200× lower cost than enterprise AML tools.
The Problem
AML compliance is a legal obligation — but enterprise tools are priced out of reach for most banks
Compliance is Legally Mandatory
Every bank processing SWIFT messages must screen for money laundering, sanctions violations, and financial crime. Failure means criminal liability — not just fines.
$500K–$5M/yr Enterprise Price Tag
Actimize, FCCM, and Fircosoft cost $500K–$5M per year. Mid-tier and regional banks legally need compliance but simply cannot afford these platforms.
Excel-Based Manual Screening
Most mid-tier banks currently run AML compliance on Excel spreadsheets and manual officer review — SWIFT messages go unscreened or under-screened.
Data Residency Violations
Enterprise black-box tools send transaction data to external servers, creating GDPR and MAS TRM data residency violations that regulators increasingly scrutinise.
No Explainable Audit Trail
Existing tools produce machine-formatted output, not human-readable audit narratives. Compliance officers and regulators cannot understand or defend flagging decisions.
Novel Patterns Go Undetected
Rule-based systems only catch known patterns. Emerging structuring tactics and new shell-company profiles bypass deterministic rules entirely.
The Solution
Enterprise-grade AML compliance on an open-source stack — at a fraction of the cost
Before: Excel & Manual Review
Mid-tier banks screen SWIFT messages manually using Excel spreadsheets and compliance officers. Transactions go unscreened. Sanctions violations are missed. A single regulatory failure means billion-dollar fines and criminal liability for compliance officers — not a slap on the wrist.
After: 4-Tier AI Decision Chain
SwiftIQ Guard screens every SWIFT message through four complementary analysis tiers in under 3 seconds, produces an AI-generated audit narrative in plain English, and stores a tamper-proof record — all running entirely within the institution's own network at $500–$3,000/month infrastructure cost.
Annual licensing for Actimize, FCCM, Fircosoft — pricing most banks out
Per month infrastructure cost — 200× cheaper, fully on-premise
The 4-Tier AI Decision Chain
Every SWIFT message passes all four tiers. Safe-Path consensus: the highest severity always wins.
Blocks known-bad patterns before any AI is involved — zero latency, zero cost per transaction.
- High-Risk Corridor Detection against FATF blacklist countries
- Structuring Alerts: amounts $9,500–$9,999 near the $10,000 CTR threshold
- TBML Scrutiny: dual-use HS codes in MT700 trade finance messages
- CRITICAL flag immediately activates Safe-Path override
Catches novel money-laundering patterns that have never been defined in rules — trained on 50K real transactions.
- XGBoost four-class risk classification (LOW / MEDIUM / HIGH / CRITICAL)
- Isolation Forest unsupervised anomaly detection for unusual profiles
- Six feature vectors: amount, corridor, round-number, threshold proximity, account age, 7-day frequency
- Flask microservice — independently scalable and upgradeable
Screens both parties against 3 global watchlists simultaneously — catches spelling variants that rule-exact-match would miss.
- Screens :50: (originator) and :59: (beneficiary) fields
- OFAC SDN, EU Consolidated, and UN Security Council lists in parallel
- Levenshtein fuzzy match at 80% threshold catches deliberate typos
- "TEHERAN IMPORT EXPORT CO" detected at 92% similarity vs. OFAC list
Produces a regulator-ready audit narrative — in plain English — without any data leaving the bank's network.
- Safe-Path consensus: highest severity from any tier wins (no averaging)
- On-premise Llama 3.1 via Ollama — zero data sent to external APIs
- 2–4 sentence professional audit narrative at temperature 0.2 for consistency
- Stored in Supabase with INSERT-only RLS — even DB admins cannot delete
Operational Workflow
From raw SWIFT message to regulator-ready audit record — in under 3 seconds
Market Positioning
A distinct value proposition for every institution tier
AML is legally required. No budget for enterprise tools. Currently on Excel. SwiftIQ Guard is the complete standalone, affordable platform.
Adds the narrative + explainability layer that Actimize and FCCM cannot provide — ISO 20022 gap flagging and challenger scoring.
SWIFT screening required but no compliance team. SwiftIQ delivers compliance-grade screening without needing a compliance department.
Audit compliance for 10–50 small banks. White-label opportunity: deploy SwiftIQ Guard under their own brand across their entire client portfolio.
User Roles & Access Control
Three-tier RBAC enforced at every API route — ensuring segregation of duties
Full system access. Configures risk thresholds, manages users, screens messages, and exports full audit logs for regulatory reporting.
Day-to-day operational role. Screens messages and takes actions on compliance queue — clear, block, or escalate. Cannot configure the system.
Read-only access for internal auditors, regulators, and senior management. Satisfies regulatory requirements for independent oversight.
Technology Stack
Open-source. On-Premise. No vendor lock-in.
4-Layer Test Coverage
Every tier validated independently and in combination — 105/105 tests passing
Parsers, rule engine (structuring, TBML, corridors), fuzzy matcher, Safe-Path consensus, RBAC permissions.
Flask ML endpoints, full 4-tier pipeline, Supabase case persistence, audit log immutability, API RBAC enforcement.
Auth flows, screening UI, queue case actions, audit log access by role — across 3 browsers.
Pipeline P95 < 3s, ML microservice P95 < 500ms, error rate < 1% under 100 concurrent users.
Key Engineering Decisions
Four architectural choices that define the platform's reliability and regulatory defensibility
On-Premise LLM is Not a Compromise — It Is a Feature
Sending transaction data to OpenAI or Anthropic APIs would violate banking data residency policies under GDPR and MAS TRM. Running Llama 3.1 locally delivers audit narrative quality indistinguishable from frontier APIs for this structured task — with no external rate limits and predictable latency.
The Safe-Path Policy Eliminates a Category of False Negatives
Early versions used weighted averages. A CRITICAL TBML flag could be diluted by a LOW ML score, passing a danger transaction to manual review. The maximum-severity Safe-Path policy eliminates this failure mode entirely — asymmetric consequences demand asymmetric policy.
Fuzzy Match Threshold is a Business Decision, Not a Technical One
The 80% Levenshtein threshold was set in consultation with compliance requirements — not optimised for F1 score alone. Too low floods the compliance queue. Too high lets deliberate phonetic variants of sanctioned names pass undetected.
Immutability Enforced at the Database Level, Not the Application
Supabase row-level security enforces INSERT-only on the audit log. A direct SQL DELETE returns 403 even for database administrators. This is the guarantee regulators need and lawyers can defend in court.
Product Roadmap
Planned enhancements to deepen coverage and automate compliance further
Auto-update OFAC/EU/UN lists daily via published APIs. Firefox E2E coverage in CI pipeline.
Front-end for batch file submission. FATF dynamic country risk scoring matrix with 200 countries.
camt.053, pacs.008, pacs.004 message types. Automated monthly XGBoost retraining on flagged cases.
Direct SWIFT Alliance Gateway — messages screen automatically as received. Correspondent bank risk profiles.
Ready to Modernise Your AML Compliance?
Let us deploy SwiftIQ Guard for your institution — enterprise-grade SWIFT screening, on-premise AI narrative generation, and a tamper-proof audit trail at a fraction of the cost of traditional compliance platforms.