FinSight AI is a financial transaction fraud investigation assistant that helps analysts prioritize suspicious transactions. It does not automatically declare a transaction as fraud. Instead, it combines machine learning, rule-based risk analysis, regulatory guidance, and similar historical cases to generate an investigation report that supports analyst decision-making.

---

Features

• Transaction risk analysis using an Isolation Forest model
• 19-feature transaction risk assessment pipeline
• Explainable rule-based risk engine
• Severity scoring from 0–100
• Local RAG-style knowledge retrieval using TF-IDF and cosine similarity
• Regulatory references from RBI, FATF, and FinCEN
• Historical fraud case retrieval
• Investigation reports with risk indicators and recommended analyst actions
• Regulatory knowledge base search
• Light/Dark mode dashboard

---

Tech Stack

Backend

• FastAPI
• Python
• scikit-learn
• Pandas
• NumPy

Machine Learning

• Isolation Forest
• PaySim Dataset (6.36M synthetic financial transactions)

Knowledge Retrieval

• TF-IDF
• Cosine Similarity

Frontend

• HTML
• CSS
• JavaScript

---

System Workflow

User Transaction
       │
       ▼
Feature Engineering (19 Features)
       │
       ▼
Isolation Forest
       │
       ├── Normal
       │
       └── Suspicious + Anomaly Score
                │
                ▼
Rule-Based Risk Engine
                │
                ▼
Severity Score (0–100)
                │
                ▼
Knowledge Retrieval
(TF-IDF + Cosine Similarity)
                │
                ▼
RBI / FATF / FinCEN Guidance
+ Historical Fraud Cases
                │
                ▼
Investigation Report
                │
                ▼
Analyst Review

---

Feature Engineering

Each transaction is transformed into the same feature format used during model training.

Features include:

• Transaction type
• Transaction amount
• Sender balance before transaction
• Sender balance after transaction
• Receiver balance before transaction
• Receiver balance after transaction
• Sender account drained flag
• Receiver balance update failure
• Sender accounting mismatch
• Receiver accounting mismatch
• Amount relative to sender balance
• High-value transaction flag (> ₹200,000)
• Merchant destination flag
• Log-transformed amount
• Hour of transaction
• Day information
• Additional engineered numerical features used during model training

---

Machine Learning Model

FinSight AI uses an Isolation Forest, an unsupervised anomaly detection algorithm.

Dataset

• PaySim synthetic financial transaction dataset
• Approximately 6.36 million transactions

Model Output

• Normal / Suspicious
• Anomaly Score

The model learns normal transaction behavior and identifies unusual transactions.

Since Isolation Forest is an unsupervised algorithm, fraud labels were not used during training. Fraud labels were used only for evaluation.

Model Performance

• ROC-AUC: 0.9112

The anomaly score is used as a triage signal, not a final fraud decision.

---

Rule-Based Risk Engine

The rule engine evaluates observable fraud indicators including:

• Account draining
• High-value transfers
• Balance inconsistencies
• Receiver balance update failures
• Nearly complete balance transfers
• TRANSFER and CASH_OUT transaction types
• Off-hours activity (1 AM–5 AM)
• Machine learning anomaly result

The total risk points are converted into a severity score.

Score	Severity
0–24	LOW
25–44	MEDIUM
45–69	HIGH
70+	CRITICAL

---

Knowledge Retrieval

FinSight AI uses a local RAG-style retrieval system.

Instead of vector embeddings and an LLM, the project performs document retrieval using:

• TF-IDF
• Cosine Similarity

The knowledge base contains:

• RBI fraud guidance
• FATF recommendations
• FinCEN fraud typologies
• Historical fraud investigation cases

Retrieved documents are used to generate deterministic investigation reports.

---

Investigation Report

For suspicious transactions, the system generates a structured report containing:

• Executive Summary
• Triggered Risk Indicators
• Regulatory References
• Similar Historical Cases
• Recommended Analyst Actions

Depending on the calculated severity, suggested actions may include:

• Enhanced monitoring
• Beneficiary verification
• Temporary hold on destination account
• Sender re-verification
• Device/IP investigation
• SIM swap verification
• SAR consideration

---

Dashboard

The web dashboard provides:

• Investigation Desk
• Transaction input form
• Detailed investigation results
• Regulatory Knowledge Base search
• Light/Dark theme support