LIVE  ·  Problem Statement 3 · X-Bank of India

Next-Generation
AML Intelligence

Graph Neural Networks that trace every rupee through the financial system — detecting laundering before it lands.

93%Precision
12msInference
5AML Typologies
LineMVGNN v2.0
Processing
Memgraph · 5,000 accounts · 500K transactions
0%

Model

Precision

0ms

Inference

Latency

0

AML

Typologies

FIU-IND

FIU-IND

ARF/TRF XML

GNN Pattern Detection

What we detect

Five money-laundering patterns identified by GNN graph traversal and pattern-aware encoding

Rapid Layering

Funds moved through accounts in minutes. Traced via velocity burst detection and temporal edge analysis.

velocity_24h feature

Circular Round-Trip

A→B→C→A round-trip flows detected via Cypher path matching (3–6 hop cycles) in the property graph.

MATCH p=(a)-[:TXN*3..6]->(a)
₹10L

Structuring / Smurfing

Transactions clustered ₹9,000–9,999 to stay below ₹10 lakh reporting threshold. Automated cluster detection.

StructuringDetector

Dormant Activation

180+ day inactive accounts suddenly receiving large transfers. GraphSAGE inductive inference on unseen nodes.

GraphSAGE zero-shot

KYC–Transaction Mismatch

Customer risk profile vs. actual behavioral pattern discrepancy — caught at edge classification layer.

PatternAwareEncoder

System Design

System architecture

Four layers of intelligence, end to end

DATA LAYER

Memgraph

›500K+ transactions
›Cypher queries
›Property graph
ML INFERENCE

LineMVGNN v2.0

›GraphSAGE
›PatternEncoder
›FP16 precision
FASTAPI BACKEND

REST API

›17 endpoints
›Prometheus metrics
›Real-time monitor
DASHBOARD

Next.js Frontend

›Pyvis graphs
›SAR XML gen
›Alert triage

~12ms

inference latency

8 GB

RTX 4050 VRAM

FP16

mixed precision

Technology

Built on

Graph
Memgraph
GQLAlchemy
Cypher
ML
PyTorch
PyG
LineMVGNN
GraphSAGE
Backend
FastAPI
uvicorn
Pydantic
Prometheus
AI
Gemini 2.0 Flash
LangChain
FIU-IND XML

Pipeline

The intelligence pipeline

01

Ingest raw transactions

AMLSim CSV data ingested into Memgraph via GQLAlchemy. 500K+ transactions stored as property graph with O(1) account lookup.

$ python ingestion/ingest_amlsim.py --csv_path data/HI-Small_Trans.csv
01
02

Train LineMVGNN

Novel line-graph architecture elevates transactions to first-class entities. 50 epochs on RTX 4050 with FP16 mixed precision training.

$ python training/train.py --model line_mvgnn --epochs 50 --fp16
02
03

Detect patterns via Cypher

Real-time graph traversal queries detect circular flows (3–6 hop cycles), structuring clusters, and dormant activations.

$ MATCH p=(a:Account)-[:TRANSACTION*3..6]->(a) WHERE a.risk > 0.7 RETURN p
03
04

Generate SAR report

LangChain + Gemini 2.0 Flash auto-generates FIU-IND compliant ARF/TRF XML for direct regulatory submission.

$ chain = prompt | ChatGoogleGenerativeAI(model='gemini-2.0-flash')
04

Live Platform

See it in action

Fund flow visualization · Alert triage · SAR report generation

localhost:3000 — NextGenAML Dashboard
ACCOUNTS
ACCT-8921CRITICAL
ACCT-4512HIGH
ACCT-2247HIGH
ACCT-7703MEDIUM
ACCT-1138LOW
ACCOUNT DETAIL
ACCT-8921
CRITICAL
Fraud Prob93%
Circular0.91
Layering0.87