Google Cloud
Google Cloud Components Used
1 Cloud Run (Inference & Control Plane)
Purpose
Hosts stateless microservices for:
Fraud scoring API
Alerting logic
Integration endpoints
Why Cloud Run
Auto-scales with transaction volume
No infrastructure management
Low latency for real-time scoring
How it works
Kafka consumers (or connectors) forward enriched events to a Cloud Run endpoint
Each request contains a fully engineered feature vector
Cloud Run calls Vertex AI for scoring
2 Vertex AI (Fraud Detection Engine)
Purpose
Hosts the fraud detection model
Performs real-time inference
Supports future retraining
Model Type (MVP-appropriate)
Gradient Boosted Trees (XGBoost / AutoML)
Hybrid rules + ML scoring
Inputs A structured feature vector, e.g.:
Outputs
Vertex AI guarantees:
Consistent inference latency
Model versioning
Auditability (important for fraud systems)
3 BigQuery (Historical Analysis & Training)
Purpose
Stores historical transactions and outcomes
Enables offline model training and evaluation
Supports fraud analytics dashboards
How it works
Kafka Connect sinks data from Confluent topics into BigQuery
Used to:
Label known fraud patterns
Evaluate false positives
Retrain models
Why BigQuery
Handles massive transaction volumes
SQL-based analysis for transparency
Direct integration with Vertex AI pipelines
4 Gemini (Optional but High-Impact)
Purpose
Explains why a transaction was flagged
Improves analyst trust and UX
Example Output
“This wallet exhibited an unusually high transaction velocity combined with first-time interactions with multiple receivers, which is consistent with laundering behavior.”
Judges strongly value explainability in AI fraud systems.
Last updated