Build Real-Time Multi-Bank AML Information Sharing Agent with OpenAI Agents SDK
In response to initiatives like Singapore's COSMIC platform for collaborative information sharing to combat money laundering and terrorism financing (ML/TF), financial institutions require automated agentic triage systems. These systems must resolve complex multi-bank entity networks, identify suspicious shell company chains, and synthesize evidence for regulatory disclosures without compromising cross-border data privacy protocols.
What you are building
Build an agentic anti-money laundering triage engine using OpenAI Agents SDK and Claude 4.1 Opus to evaluate cross-bank transaction topologies and flags potential money laundering networks.
How work is evaluated
Evaluates the agent's capability to detect multi-bank layering schemes, extract entity relationships, and compute risk scores with strict schema accuracy.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
schema_conformance
Output adheres exactly to expected JSON structure.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
ring_detection_f1
F1 score for correctly identifying laundering rings. • target: 0.95 • range: 0.8-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Implement multi-bank transaction topology analysis using OpenAI Agents SDK tool execution
Perform cross-entity disambiguation and graph traversal to uncover synthetic identity rings
Structure automated SAR (Suspicious Activity Report) narratives following FATF guidelines
Integrate Claude 4.1 Opus via fallback tool definitions for complex legal reasoning
Resources
Synthetic multi-bank financial telemetry containing 10,000 accounts and 50,000 transactions with embedded circular pass-through and rapid movement anomalies.
Challenge setup
Evaluates the agent's capability to detect multi-bank layering schemes, extract entity relationships, and compute risk scores with strict schema accuracy.
Challenge input
JSON containing network graph nodes (accounts/entities) and edges (transactions).
OpenAI Agents SDK
Primary framework for tool invocation and state handling.
Claude 4.1 Opus
High-precision legal reasoning for automated SAR draft synthesis.
Evaluated output
JSON with detected_ring_id, risk_score, matched_patterns, and generated_sar_summary.
- Output adheres exactly to expected JSON structure.
- F1 score for correctly identifying laundering rings. • target: 0.95 • range: 0.8-1
- Benchmark: Financial CyberCrime Bench
- Ring Detection F1 target: 0.95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- OpenAI · Required
- Claude Code · Optional
- Claude Code · Optional
- Claude 4.1 Opus · Optional
Evaluation contract
- schema_conformance · Weight 1
- ring_detection_f1 · Weight 1
Metrics
- ring_detection_f1 · Target 0.95
Benchmarks
- Financial CyberCrime Bench
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing network graph nodes (accounts/entities) and edges (transactions).
Output: JSON with detected_ring_id, risk_score, matched_patterns, and generated_sar_summary.
Execution harness: Python
Evaluation tools
- OpenAI Agents SDK · Agent Orchestration
- Claude 4.1 Opus · Legal & Compliance LLM
Recipe freeze marker
No recipe freeze time is recorded.