Major System Components
1. Blockchain Ingestion Layer
Responsibility
Collects raw blockchain transactions in real time
Characteristics
Can be:
Blockchain node
Indexer
Transaction simulator (for MVP)
Pushes events to Kafka
Output
Immutable transaction events
2. Confluent Cloud (Streaming Backbone)
Responsibility
Durable, ordered, real-time event transport
Key Topics
raw-transactionsenriched-transactionsfraud-alerts
Why It Exists
Decouples ingestion, processing, AI scoring, and alerting
Enables replay, scaling, and fault tolerance
3. Stream Processing Layer (Confluent)
Responsibility
Real-time behavioral feature engineering
Capabilities
Sliding and tumbling windows
Wallet-level aggregation
Burst and anomaly indicators
Outputs
Feature-enriched transaction events
This layer is the core Confluent value demonstration.
4. Fraud Scoring Service (Google Cloud Run)
Responsibility
Acts as the real-time inference gateway
Characteristics
Stateless
Auto-scaling
Low-latency HTTP interface
Flow
Receives feature vectors
Calls Vertex AI
Applies policy logic
5. Vertex AI (Machine Learning Layer)
Responsibility
Computes fraud risk scores
Model Characteristics
Supervised or semi-supervised
Outputs probabilistic risk score
Versioned and auditable
Why Vertex AI
Managed inference
Model lifecycle control
Native integration with BigQuery
6. Decision & Policy Layer
Responsibility
Converts risk scores into actions
Example Policies
Allow
Flag for review
Escalate / block (off-chain)
Key Point
Decisions are configurable and explainable
No hard coupling to on-chain execution
7. Analytics, Audit & Explainability
BigQuery
Stores:
Transactions
Features
Risk decisions
Used for:
Model retraining
Compliance
Historical analysis
Optional: Gemini
Produces human-readable explanations
Improves trust and analyst productivity
Last updated