Fraud Detection Data Pipeline

Build steps

1. Generate transactions, including injected suspicious patterns
2. Compute windowed features per card with event-time windows
3. Join features to each transaction
4. Apply documented rules (for example many transactions in a short window, or unusually high amounts relative to the card's history)
5. Write flagged transactions with the rule that fired to a review table
6. Backtest rules on stored raw data and measure how many injected patterns they catch
7. Document limitations

Testing

• Unit tests for each rule with crafted inputs
• Backtest on generated data with known injected patterns

Data quality

• Features non-null after warm-up
• Every flag records the rule and feature values that triggered it

Monitoring

• Flag rate over time (sudden changes indicate data or rule problems)
• Streaming latency and lag

Cost considerations

• Runs locally
• State size grows with the number of active cards and window lengths

Interview questions this project prepares you for

• Why compute features in streaming rather than batch?
• How do you avoid using future information when backtesting?
• How do you explain why a transaction was flagged?
• How would a machine-learning model use your feature table?

Resume bullets (add only what you measured)

Use these as structure. Replace the wording with results you actually produced and can show; never claim numbers you did not measure.

• Built a streaming fraud-detection data pipeline computing per-card rolling features and explainable rule flags from a Kafka transaction stream
• Backtested rules on stored synthetic data with injected patterns and documented detection rate and limitations

Extensions

• Train a simple model on the feature table and compare it with rules
• Add feedback from reviewers to label data