AML Sentinel
China AML Intelligence Platform
End-to-end AML compliance workbench: PBOC threshold alerts → network graph investigation → regulatory RAG Q&A → AI Agent auto-drafts STRs in CAMLMAC format. Fully simulates a China AML compliance analyst workflow.
PBOC threshold rules + DGraph-Fin ML scores flag 500 real transactions across 6 rule types
Lifecycle tracking, account relationship network, ECharts force graph with typology detection
BM25 client-side retrieval over 55 chunks from 4 real AML docs → claude-sonnet-4-6 streaming with article citations
6-step claude-sonnet-4-6 tool_use loop: transaction extraction → account analysis → regulation retrieval → typology classification → urgency assessment → CAMLMAC format report
Full Pipeline: Data to Compliance Decision
HI-Small CSV → networkx subgraph → 50-node force graph
PyG 20-dim features → 500-row sample → PBOC rule scoring
AML Law/FATF/KPMG/CAMLMAC → clause-level chunking → 55 chunks
src/app/api/aml/rag/route.tsServer RoutePOST → BM25 top-5 chunks → claude-sonnet-4-6 streaming → SSE cited answer
src/app/api/aml/str/route.tsServer Route6-iteration tool_use loop → per-step SSE events → final CAMLMAC STR stream
RAG answer gen + STR tool_use main loop
Server-to-client real-time streaming, zero-latency UX
Auto-fallback to deterministic mock when API key absent
Aligned with FinTech AI Product JD Requirements
BM25 keyword retrieval + claude-sonnet-4-6 streaming with per-claim article citations
claude-sonnet-4-6 · BM25 · SSE6-step tool_use loop: 5 deterministic tools driven autonomously by Claude
tool_use · agentic loopPBOC reporting thresholds + IBM AML 12 typologies + CAMLMAC STR format
PBOC §31/§33 · FATF Rec.16 · CAMLMACDGraph-Fin ML precomputed scores + 6 PBOC rules real-time trigger + 4-tier risk tiers
DGraph-Fin · NeurIPS 2022 · 4-tierAML Law Art.31/33 → filing deadlines; FATF IO5 → beneficial ownership detection
反洗钱法 2024 · FATF IO5/IO6Static TypeScript data layer, no DB; Vercel edge deploy; SSE streaming API
Next.js 14 · Vercel · no-DB