Inspiration
India’s digital payment systems have made money movement instant — but fraud has evolved just as fast.
While studying UPI fraud cases, we observed that many victims were not hacked technically, but were pressured into acting through urgency, fear, and social engineering using mule accounts.
Most existing fraud systems focus on transaction amounts and static rules, but ignore human behavior, intent, and network relationships.
This inspired us to ask a different question:
Can we stop fraud before money moves by understanding behavior, context, and relationships — not just transactions?
This question led to the creation of CHRONOS-BHARAT.
What it does
CHRONOS-BHARAT is a real-time, pre-transaction fraud prevention system for digital payments.
Instead of analyzing transactions after completion, it:
- Intercepts payment requests before settlement
- Evaluates risk using temporal patterns, behavioral signals, geography, and transaction notes
- Detects mule-network behavior through graph analysis
- Blocks suspicious payments instantly
- Explains why a transaction was flagged in clear, human-readable terms
The system simulates a real-world flow where a payment is either approved or blocked before money moves.
How we built it
We built CHRONOS-BHARAT as a modular and explainable AI pipeline.
Dataset Design & Enrichment
Since real bank data is unavailable, we used synthetic transaction data and enriched it with India-specific context:
- Indian pincode and latitude-longitude data for geographic behavior
- Festival calendar data (Diwali, Eid, etc.) to normalize transaction bursts
- Hinglish transaction notes to model realistic payment behavior and scam language
Fraud Signal Engine
Each transaction is evaluated using multiple signals:
- Transaction amount anomalies
- Suspicious timing (e.g., late-night activity)
- Coercive or high-pressure transaction notes
- Burst frequency patterns
- Network behavior such as one-to-many transfers
No single signal decides fraud — the system uses defense in depth.
Network & Temporal Analysis
Transactions are modeled as a temporal graph:
- Nodes represent accounts
- Edges represent money flow over time
This allows detection of organized mule networks instead of isolated anomalies.
Explainable Decision Layer
For every flagged transaction, the system generates:
- A risk score
- Top contributing factors
- Clear explanations suitable for banks and regulators
Demo Interface
A split-screen demo shows:
- A simulated payment app (left)
- A live fraud detection dashboard (right)
This allows judges to see fraud prevention happening in real time.
Challenges we ran into
- No access to real banking data, requiring careful synthetic data design
- Avoiding false positives for legitimate urgent or high-value payments
- Balancing realism with hackathon-level simplicity
- Designing explanations that are understandable to non-technical stakeholders
Accomplishments that we're proud of
- Demonstrated fraud prevention before transaction settlement
- Built an explainable system rather than a black-box model
- Designed India-specific fraud signals instead of generic features
- Created a clear, real-time demo suitable for judges
- Used only compliance-safe synthetic data
What we learned
- Fraud detection is a human behavior problem as much as a technical one
- Context such as language, time, and culture is critical
- Explainability is essential for financial systems
- Graph-based analysis reveals patterns tabular data cannot
- A clear demo is more impactful than complex but opaque models
What's next for CHRONOS-BHARAT
- Adding agent-based automated fraud investigation reports
- Expanding to cross-bank and cross-platform fraud graphs
- Introducing adaptive user interventions instead of only blocking
- Exploring privacy-preserving learning techniques
- Preparing the system for regulator-grade evaluation
Log in or sign up for Devpost to join the conversation.