Building an Agentic AML Risk Intelligence Workflow with OpenAI Agents SDK and Galileo
In light of MAS guidelines on agentic AI in financial services and the Collaborative Sharing of ML/TF Information & Cases (COSMIC) platform, financial institutions in Singapore are adopting autonomous multi-agent systems to detect money laundering and terrorist financing across inter-bank networks. Regulated financial agents must dynamically inspect complex multi-hop transaction graphs, verify shell company ownership patterns, evaluate cross-border fund flows, and output explainable audit traces while adhering to strict compliance constraints. In this challenge, you will construct a production-ready anti-money laundering (AML) intelligence system using the OpenAI Agents SDK to manage multi-agent orchestration and tool invocation. To ensure continuous governance and strict risk bounds, you will integrate Galileo to evaluate output accuracy, detect hallucinations in risk summaries, and enforce safety guardrails before regulatory handoff. Your solution will consume real-time synthetic banking transaction feeds, query COSMIC inter-bank flag databases via function tools, dynamically isolate suspicious networks, and generate fully structured Suspicious Transaction Report (STR) dossiers complete with human-in-the-loop review triggers.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop a multi-agent AML risk analysis pipeline utilizing the OpenAI Agents SDK for inter-bank transaction investigation and Galileo for regulatory output evaluation and guardrailing.
How work is evaluated
Evaluates the agent system on correctly identifying AML anomalies, tool usage compliance using OpenAI Agents SDK, and output guardrailing verified by Galileo metrics.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Flag Detection Accuracy
Agent must correctly identify all layered money laundering accounts linked to COSMIC flags.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Galileo Hallucination Guardrail Check
Galileo evaluation score for hallucination index must remain below 0.05.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
detection_f1_score
F1 score for accurately flagging illicit accounts and transaction nodes • target: 0.95 • range: 0.85-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
galileo_context_adherence
Galileo metric score for factual grounding in regulatory inputs • target: 0.98 • range: 0.9-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Orchestrate multi-agent financial investigations using the OpenAI Agents SDK and function calling.
Integrate Galileo evaluation metrics to continuously measure ground-truth context adherence and hallucination rates in generated risk dossiers.
Implement automated transaction graph traversal tools to spot multi-hop layered money laundering patterns.
Build regulator-friendly audit trails and structured human-in-the-loop review triggers for high-risk flags.
Reference links and supporting material
Synthetic multi-bank transaction ledger containing 1,000 corporate account records, transaction edge histories, and ground-truth COSMIC high-risk entity flags.
Challenge setup
Evaluates the agent system on correctly identifying AML anomalies, tool usage compliance using OpenAI Agents SDK, and output guardrailing verified by Galileo metrics.
Challenge input
JSON object containing bank_transactions array, cosmic_flagged_entities list, and customer_profiles array.
OpenAI Agents SDK
Required for orchestrating multi-agent compliance workflows and handoffs.
Galileo
Required for evaluation, hallucination detection, and audit guardrails.
Python
Core environment for running the agent harness and processing transaction data.
Evaluated output
JSON object with risk_assessments array, suspicious_flagged_accounts array, str_dossiers list, and galileo_eval_summary.
- Agent must correctly identify all layered money laundering accounts linked to COSMIC flags.
- Galileo evaluation score for hallucination index must remain below 0.05.
- F1 score for accurately flagging illicit accounts and transaction nodes • target: 0.95 • range: 0.85-1
- Benchmark: MAS FinTech AI Agent Benchmark 2026
- Detection F1 Score target: 0.95
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- OpenAI · Optional
- RAI · Optional
- Galileo · Required
- Galileo · Required
Evaluation contract
- Flag Detection Accuracy · Weight 1
- Galileo Hallucination Guardrail Check · Weight 1
- detection_f1_score · Weight 1
- galileo_context_adherence · Weight 1
Metrics
- detection_f1_score · Target 0.95
- galileo_context_adherence · Target 0.98
Benchmarks
- MAS FinTech AI Agent Benchmark 2026
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON object containing bank_transactions array, cosmic_flagged_entities list, and customer_profiles array.
Output: JSON object with risk_assessments array, suspicious_flagged_accounts array, str_dossiers list, and galileo_eval_summary.
Execution harness: Python
Evaluation tools
- OpenAI Agents SDK · agent_framework
- Galileo · ai_workflow_automation
- Python · programming_language
Recipe freeze marker
No recipe freeze time is recorded.