MuleNet: Identifying Suspicious Financial Mule Accounts with AI
Learn how MuleNet uses counterfactual AI, LLMs, and ensemble ML models like XGBoost to detect financial mule accounts by analyzing behavioral 'Humanity Scores'.
MuleNet
AI/ML-Based Classification of Suspicious Mule Accounts
Detect the financial life that should exist — but doesn't.
Powered by XGBoost · LightGBM · LLM · Counterfactual AI
FinCrime Track
AI Division
Team Delta
The Problem
Mule Accounts Are Hiding in Plain Sight
₹1.3T+
Lost annually to mule fraud
3 in 10
Mule accounts go undetected
48hrs
Average money transfer window
Traditional rule-based systems catch less than 40% of mule accounts
Genuine Account
Salary
Bills
ATM
Groceries
Mule Account
Bulk Inward
Rapid Outward
No lifestyle spend
CFI
Counterfactual Fraud Investigator
Detect fraud by identifying the financial life that SHOULD exist, but doesn't.
Bank Data
Raw transactions & events
Feature Engineering
Behavioral signals
Persona Engine
Customer clustering
Counterfactual Generator
Expected baseline
Humanity Score
Deviation metric
Fraud ML Model
Risk probability
LLM Investigator
Contextual analysis
Analyst Dashboard
Actionable insights
CFI doesn't just score accounts — it reconstructs what a genuine financial life should look like, then measures the gap.
MODULE 01–02
Raw Data → Intelligent Features
Feature Categories
Transaction Features
Avg amount, Median, Max/Min amount
Velocity Features
Txns/day, Inward-Outward gap, Night ratio
Diversity Features
Unique beneficiaries & senders
Stability Features
Monthly periodicity score
Input Specification
F115 → F3894
Target: F3924
Transaction
35%
Velocity
28%
Diversity
22%
Stability
15%
3,780 Features → Engineered to ~25 High-Signal Indicators
MODULE 02–03
Who is This Person?
Persona Engine
Salaried
Salary · Bills · ATM · Groceries · Fuel
Student
Recharge · Food · UPI transfers · Low balance
Business
Suppliers · B2B transfers · Invoice payments
Retired
Pension · Medical · Low velocity · Regular SIP
Counterfactual Generator
IF Genuine (Salaried), EXPECT:
Salary
Utility Bills
Merchant Spend
ATM Withdrawals
Trained ONLY on genuine accounts (F3924=0) to learn authentic spending DNA
The Gap That Reveals Everything
MODULE 04–05
Reality Comparison Engine
Humanity Score Gauge
Expected Behavior
Category
Observed Reality
Salary credit monthly
INCOME
No regular income ❌
Utility bills paid
BILLS
No bills detected ❌
ATM withdrawals
CASH
No ATM usage ❌
Grocery merchants
SPEND
No retail spending ❌
Fuel transactions
MOBILITY
No fuel/travel ❌
18%
POTENTIAL MULE
Normal (80-100)
Slightly Abnormal (60-80)
Suspicious (40-60)
Highly Suspicious (20-40)
⚠️ Potential Mule (<20)
Humanity Score = (Observed / Expected) × 100
A genuine human leaves financial footprints. Mules don't.
Triple-Layer ML Ensemble
MODULE 06
XGBoost
Gradient boosted trees, captures non-linear feature interactions
High Precision
Random Forest
Ensemble of decision trees with bootstrapping, robust to outliers
Low Variance
LightGBM
Leaf-wise tree growth, handles large feature space efficiently
Fast & Scalable
Original Features (F115-F3894)
Engineered Features (~25)
Humanity Score
Fraud Probability Score
Stacked Ensemble: Meta-learner combines all three model outputs for final classification
MODULE 07–08
From Score to Story
LLM Investigator
Salaried Employee
Salary credit, Bills, ATM, Groceries, Fuel
Bulk inward transfers × 3, Immediate dispersal to 8 accounts, No lifestyle spending
Risk Score: 92% | Humanity Score: 18%
⚠️ INVESTIGATION SUMMARY: Account exhibits classic mule behavior — large inward transfers immediately dispersed to multiple beneficiaries with zero lifestyle spending. Recommend immediate freeze and SAR filing.
Analyst Dashboard — One Screen, Full Picture
92%
18%
Salaried
Expected Life
Observed Life
AI Investigation Report
Account flagged for zero retail footprint. Immediate dispersal directs funds entirely to outbound typologies. System recommends auto-freeze and SAR generation.
FREEZE ACCOUNT
ARCHITECTURE
End-to-End System Flow
Bank Transaction Dataset
Feature Engineering Pipeline
Genuine Account Filter (F3924=0)
KMeans/GMM Clustering
Persona Assignment Engine
Counterfactual Generator
Reality Comparison Engine
Humanity Score Calculator
Fraud ML Classifier (XGBoost+RF+LGBM)
LLM Investigator (GPT-4/Gemini)
Analyst Dashboard + Alert System
OUR EDGE
Why CFI Wins the Hackathon
Most Teams
Dataset
XGBoost
Fraud Score
What's missing?
No explainability
No account profiling
Black box prediction
No narrative for analyst
CFI
Dataset
Who is this person?
What life should exist?
What life actually exists?
Missing Humanity
Fraud Probability
What we deliver:
Persona-based counterfactuals
Explainable Humanity Score
LLM investigation narrative
Analyst-ready dashboard
Interpretable AI
Novel Counterfactual Approach
LLM Integration
End-to-End Pipeline
We don't just detect fraud. We reconstruct the truth.
MuleNet
CFI — Counterfactual Fraud Investigator
Detect fraud by identifying the financial life that should exist, but doesn't.
8 Modules
Complete Pipeline
3 ML Models
Ensemble Approach
1 Goal
Stop Financial Crime
Tech Stack
Python
XGBoost
LightGBM
Random Forest
KMeans
GMM
LLM/GPT
Streamlit Dashboard
Questions?
Built for Hackathon Selection 2026
- fintech
- fraud-detection
- ai-ml
- fincrime
- counterfactual-ai
- machine-learning
- xgboost
- llm