Build an Agentic AML Entity Resolution Engine with Claude Agents SDK
With regulatory frameworks like Singapore MAS COSMIC and global financial crime bodies pushing for collaborative, autonomous intelligence sharing, financial institutions require next-generation Anti-Money Laundering (AML) triage systems. Traditional rules-based transaction monitoring generates overwhelming false-positive alerts, slowing down human investigators and leaving complex multi-institutional laundering rings undetected. In this challenge, you will design and implement an autonomous AML Case Triage and Entity Resolution agent utilizing the Claude Agents SDK powered by Claude 4 Opus. The system will process streaming transaction alerts, resolve noisy multi-bank entity identifiers across synthetic graph data, execute extended thinking loops for risk scoring, and produce explainable audit trails compliant with regulatory expectations. By leveraging the extended thinking, tool call orchestration, and structured output abilities of the Claude Agents SDK, your solution will autonomously cross-reference entity risk profiles, score money laundering typologies (such as structuring, layer-mule rings, and rapid velocity movement), and generate formatted SAR (Suspicious Activity Report) draft briefs with minimal human intervention.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an autonomous AML entity resolution and triage agent using Claude Agents SDK and Claude 4 Opus to process graph transaction streams and detect financial crime networks.
How work is evaluated
Evaluates the agent's ability to accurately resolve merged/duplicate entities, correctly detect layering typologies, and generate compliant SAR risk scores with explanations.
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.
schema_conformance
Output adheres strictly to expected JSON structure.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
high_risk_flagging
Correctly flags known layering case as sar_recommended=true.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
entity_resolution_f1
F1 score of entity alias resolution matching gold canonical mappings. • target: 0.85 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
risk_scoring_mae
Mean Absolute Error compared to expert baseline risk scores. • target: 0.1 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Implement multi-tool agentic workflows using the Claude Agents SDK for complex financial crime investigation.
Execute graph-based entity resolution across disparate transaction records with high accuracy.
Construct explainable risk scoring algorithms for financial laundering typologies.
Deploy deterministic audit trail generation and automated SAR drafting adhering to regulatory standards.
Reference links and supporting material
Synthetic multi-bank financial transaction dataset containing 1,000 noisy entity accounts, fuzzy name variations, transaction timestamped flows, and labeled money laundering ground truth topologies.
Challenge setup
Evaluates the agent's ability to accurately resolve merged/duplicate entities, correctly detect layering typologies, and generate compliant SAR risk scores with explanations.
Challenge input
JSON containing noisy transaction records, account profiles, and entity aliases
Claude Agents SDK
Required framework for building autonomous agentic loops with extended reasoning.
Claude 4 Opus
Primary LLM model offering deep reasoning and structured tool execution capabilities.
Evaluated output
JSON containing resolved entity mappings, risk score, typology classification, and investigation rationale
- Output adheres strictly to expected JSON structure.
- Correctly flags known layering case as sar_recommended=true.
- F1 score of entity alias resolution matching gold canonical mappings. • target: 0.85 • range: 0-1
- Benchmark: Financial Crime Detection Benchmark (FinCrimeBench)
- Entity Resolution F1 target: 0.85
- 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
- Claude Code · Required
- Claude Code · Required
- Claude 4.1 Opus · Optional
- Claude 4 Sonnet · Optional
Evaluation contract
- schema_conformance · Weight 1
- high_risk_flagging · Weight 1
- entity_resolution_f1 · Weight 1
- risk_scoring_mae · Weight 1
Metrics
- entity_resolution_f1 · Target 0.85
- risk_scoring_mae · Target 0.1
Benchmarks
- Financial Crime Detection Benchmark (FinCrimeBench)
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing noisy transaction records, account profiles, and entity aliases
Output: JSON containing resolved entity mappings, risk score, typology classification, and investigation rationale
Execution harness: Python
Evaluation tools
- Claude Agents SDK · agent_framework
- Claude 4 Opus · llm_engine
Recipe freeze marker
No recipe freeze time is recorded.