AML Sentinel: China AML Investigation Workbench
A bank-style AML compliance workbench combining real Chinese financial fraud data (DGraph-Fin by Finvolution), PBOC regulatory RAG, and an AI Agent that auto-drafts Suspicious Transaction Reports — designed around the full analyst investigation workflow.
Overview
Key Features
Alert Queue
PBOC threshold rules applied to transaction data; severity triage (Critical / High / Medium / Low) with rule explanations and risk score breakdowns from DGraph-Fin ML scores
Case Overview
Structured case lifecycle (6 states), key findings, reporting urgency countdown, analyst notes with handoff to supervisor review
Investigation Workspace
ECharts force-directed graph with fan-out, fan-in, and cycle pattern highlighting from IBM AML HI-Small typology labels; investigation checklist
Regulation Copilot
Hybrid BM25 + Anthropic API Q&A over 反洗钱法 2024, FATF China MER 2019, and KPMG AML analysis; every answer cites specific article or section references
STR Auto-Draft
6-step AI Agent using claude-sonnet-4-6 tool_use: case extraction, account analysis, regulation retrieval, typology classification, urgency assessment, CAMLMAC-format report generation
Methodology
Data layer: DGraph-Fin node features mapped to alert risk signals; IBM AML HI-Small provides labeled laundering typology patterns for the network graph. Regulatory RAG uses CJK-aware BM25 retrieval to fetch top-5 chunks from ~80 pre-chunked regulatory excerpts, then passes them to Anthropic claude-sonnet-4-6 via streaming API for grounded answer generation. STR generation uses an agentic tool_use loop: Claude iteratively calls 5 deterministic tools, then produces a structured STR draft in CAMLMAC format. The system always maintains human-in-the-loop: AI output is advisory, never auto-submitted.
Tech Stack
Project Info
Launch AML Sentinel Workbench
Alert triage → Case investigation → Network graph → Regulatory RAG → AI Agent STR drafting. Full product experience.