Hackathon Problem
Statement
TABLE OF CONTENTS
• Context
• What you need to Build
• Fraud Detection API: Real time
• Fraud Detection API: Batch
• Fraud Reporting API
• Transaction and Fraud Monitoring Dashboard
• What will be provided to you
• How will you be evaluated
Context
We are a Payment Gateway. We receive a payment transaction request from a payer to
a payee. We authorise or reject the detection based on various parameters like enough
funds in the payer account, approval from payer’s bank, etc.
As part of the authorisation we also need to check if the transaction is fraudulent or not.
Now labeling a transaction as fraudulent is not just a rule based deterministic process, it
is a probabilistic dynamic process.
And this is where AI comes in. We want to combine the power of expert rules with that
of AI models to detect frauds better.
We also want to monitor how our detection is faring against actually reported errors
What you need to Build
Build a Fraud Detection, Alert, and Monitoring (FDAM) system
The FDAM system comprises the below mentioned components. Build all of them.
Fraud Detection API: Real time
• Input: A single transaction data in JSON format in the request body
• Output:
“transaction_id”: <string>,
“is_fraud”: <boolean>,
“fraud_source”: <string: “rule”/”model”>,
“fraud_reason”: <string>,
“fraud_score”: <float>
}
• It should have a front end to configure the rules of the rule engine
• It should have an AI model to predict is_fraud
• The model can output just a boolean/binary value, or a score which then can be converted into a binary decision. In
case of latter return the score in fraud_score field
• It should store all the input and output data to an external database in the table named fraud_detection with field
is_fraud recorded as is_fraud_predicted
• It should be real time with average latency < 300 ms
Fraud Detection API: Batch
• Input: Data of multiple transactions in JSON format in the request body
• It should be using the Fraud Detection Real Time API in the backend
• Its should process multiple transactions in parallel but latency can vary depending on underlying server’s
memory and compute
• Output:
“<transaction id>”: {
“is_fraud”: <boolean>,
“fraud_reason”: <string>,
“fraud_score”: <float>
},
...
}
Fraud Reporting API
• Input:
“transaction_id”: <string>,
“reporting_entity_id”: <string>“,
“fraud_details”: <string>,
• It too should store all the input and output data to the same external database as used by the Fraud Detection
API in the table named fraud_reporting with field a is_fraud_reported set to True
• Output:
“transaction_id”: <string>,
“reporting_acknowledged”: <boolean>“,
“failure_code”: <int>
}
Transaction and Fraud Monitoring
Dashboard
It should have the following
• A table to show raw transaction data including is_fraud_predicted and is_fraud_reported columns
• Data should be filterable by
• Dates
• Payer ID
• Payee ID
• Data should be searchable by
• Transaction ID
• A dynamic graph to compare the no. of predicted and reported frauds on the following dimensions:
• Transaction Channel
• Transaction Payment Mode
• Payment Gateway Bank
• Payer ID
• Payee ID
•A time series graph to compare the trend of predicted and reported frauds
• The time frame should be selectable
• The granularity of time on x axis should dynamically vary as per the selected time frame and zoom levels
•A evaluation section showing confusion matrix, precision, and recall over a selectable time period
What will be provided to you
• A data file where each row will be a transaction. In columns you will have transaction details as received by the Fraud
Detection API as input and also a column is_fraud_reported which will be a boolean telling if the transaction was
actually a fraud or not. Below are the column names and their meaning:
• A sample request each for each of the APIs
Key Value
transaction_id Unique transaction Id
transaction_date Date on which txn is happening
transaction_amount Amount of the transaction
transaction_channel Like, web, mobile
transaction_payment_mode Card, UPI, NEFT...
payment_gateway_bank Service Bank
payer_email Sender email
payer_mobile Sender mobile
payer_card_brand Sender Card like
payer_device Device Id
payer_browser Web browser
payee_id Unique payee Id
How will you be evaluated
Deliverable Weightage
Fraud Detection API: Real Time 50%
1.Rule Engine with frontend 30%
2.AI model 40%
3.API responding or not? 10%
4.API latency 5%
5.API storing the data in a DB or not? 15%
Fraud Detection API: Batch 5%
1.API working or not? 40%
2.Is processing transactions in parallel? 60%
Fraud Reporting API 5%
Database 10%
Dashboard 30%
Raw transaction table 30%
Graph to see groupings by dimensions 20%
Time series graph 20%
Evaluation section 30%
