Corridor
Two AI agents carrying the heaviest AML workload for exchange houses — scoring and grouping transaction-monitor alerts with pre-assembled evidence, then drafting STR narratives for goAML filing under the 20-day deadline.
Project presentation
Download MP4Live system presentation
Corridor · Agent Mesh
Alert Intake
Alerts arrive from the transaction monitor — rule-based or ML — and are deduplicated.
Technical Design
Two-step AML automation for exchange houses. A triage layer scores and groups transaction-monitor alerts, assembling KYC, history, and counterparty evidence with a one-paragraph rationale per alert. Escalated cases flow into an STR drafter that produces goAML-formatted narratives for the UAE Financial Intelligence Unit, with the 20-day clock tracked end to end.
System Components
Alert Triage Service
Scores and groups transaction-monitor alerts by risk
Evidence Assembler
Collects KYC, history, counterparties, and watchlist hits
Rationale Engine
Generates one-paragraph review rationales with indicator matches
STR Drafter
Produces goAML-formatted narrative drafts with indicator codes
Deadline Tracker
Monitors the 20-day alert-to-MLRO clock
Audit Store
Immutable case files for regulator examination
Data Flow
Technology Choice
Azure AI Foundry + LangChain
Azure AI Foundry provides the regulated model hosting exchange houses need, while LangChain orchestrates the triage-to-STR agent workflow with an immutable audit trail aligned to CBUAE expectations.
Alternatives Considered
AWS Bedrock + LangGraph, Google Vertex AI + CrewAI
Core Frameworks
Azure AI Foundry
Compliant model hosting for triage and drafting
LangChain
Triage-to-STR agent orchestration
Azure AI Search
Case and evidence retrieval
FastAPI
API for the compliance dashboard
Implementation Plan
Alert Integration
- Monitor alert ingestion
- Deduplication
- Correlation keys
Triage Engine
- Risk scoring model
- Alert grouping
- False-positive filter
STR Drafting
- Narrative templates
- goAML formatting
- Indicator codes
Deadline & Audit
- 20-day clock
- MLRO sign-off
- Regulator export
Key Milestones
Risk Mitigation
Key Features
Alert Scoring & Grouping
ML scores every transaction-monitor alert and groups related alerts into one case, so the analyst reviews a single file instead of ten separate ones.
Evidence Pre-assembly
Pulls customer KYC, transaction history, counterparties, and risk factors into each alert before an analyst ever opens it.
One-Paragraph Rationale
Drafts a review-ready rationale per alert so the analyst closes benign cases or escalates in seconds instead of reading the full file.
STR Narrative Drafting
Generates the Suspicious Transaction Report narrative in the format the UAE FIU expects — who, what transactions, why suspicious, indicators matched.
goAML Formatting
Outputs draft STRs structured for the goAML portal with the correct indicator codes and case references for direct filing.
Deadline Tracking
Tracks the 20-calendar-day clock from alert to MLRO recommendation and flags cases heading toward a miss.
How It Works
Alert Intake
Alerts arrive from the transaction monitor — rule-based or ML — and are deduplicated.
Alert Scoring
Every alert gets an ML risk score from pattern, velocity, and corridor features.
Alert Grouping
Related alerts — same customer, corridor, or counterparty — are grouped into one case.
Evidence Assembly
KYC file, transaction history, counterparties, and watchlist hits are pulled per case.
Rationale Generation
AI writes a one-paragraph rationale with the indicators matched and a recommendation.
Analyst Decision
Analyst closes the case as benign with rationale, or escalates it for investigation.
STR Drafting
Escalated cases produce a goAML-formatted STR narrative with the full case package.
Filing & Deadline Tracking
MLRO reviews, verifies, and signs the STR before goAML filing, with the 20-day clock tracked.
Multi-Agent Architecture
Triage Agent
Scores and groups transaction-monitor alerts, ranking them by risk so the analyst works the queue in seconds.
- Alert scoring
- Alert grouping
- Priority ranking
- False-positive filtering
Evidence Agent
Assembles the supporting evidence pack for each alert — KYC, history, counterparties, and risk factors.
- KYC retrieval
- Transaction history
- Counterparty mapping
- Risk factor lookup
Rationale Agent
Writes the one-paragraph review rationale with matched indicators and a recommendation.
- Summary generation
- Indicator matching
- Confidence scoring
- Close / escalate call
STR Drafter
Produces the goAML-formatted STR narrative with the case package assembled and indicator codes applied.
- Narrative generation
- goAML formatting
- Indicator coding
- Case package assembly
Compliance Officer
Reviews the draft STR, verifies the facts, and signs before filing — the human stays accountable.
- Quality review
- Regulatory validation
- Digital sign-off
- Deadline management
Use Cases
Exchange House Alert Triage
Clear thousands of daily transaction-monitor alerts with grouped evidence and instant rationales.
goAML STR Filing
Draft defensible STR narratives in the UAE FIU format — deadline-safe and regulator-ready.
Corridor Risk Monitoring
Focus on the India, Pakistan, and Philippines remittance corridors regulators scrutinize most.
CBUAE Examination Readiness
Keep complete, audit-ready case files for regulator examination instead of weeks of evidence gathering.
Structuring & Threshold Detection
Catch layering and structuring just below the AED 10,000 cash reporting threshold.
MLRO Workload Relief
Free the MLRO to verify and sign drafts instead of writing STR narratives from a blank page.