Financial Fraud Detection on CSA Loom¶
Real-time credit-card transaction stream → fraud-detection ML scoring → high-risk transactions trigger alerts to investigators via Activator → analysts query historical fraud patterns via Data Agent.
What you'll build¶
Source: Cosmos DB (live transactions ledger)
↓ Loom Mirroring Engine (CDC)
Bronze: raw_transactions Delta table
↓ Databricks Spark Structured Streaming
Silver: enriched_transactions (with feature engineering for ML)
↓ Databricks ML model scoring
Gold: scored_transactions (with fraud_score 0-1)
↓ Loom Activator Engine
Real-time alert: fraud_score > 0.9 → Teams to investigators
↓ Investigator response → annotate + feedback loop
↓ Power BI: fraud-trend dashboards (Direct-Lake-Shim refresh)
↓ Loom Data Agent: NL Q&A for historical fraud patterns
Components¶
| Loom capability | Used for |
|---|---|
| Mirroring Engine | CDC from Cosmos transactions |
| Databricks notebook | Feature engineering + ML scoring |
| Databricks MLflow | Model versioning + serving |
| ADX | Real-time scored transaction stream |
| Loom Activator Engine | Threshold alert (fraud_score > 0.9 sustained) |
| Power BI semantic model | Fraud-trend BI |
| Loom Data Agent | Analyst NL Q&A |
Per-boundary notes¶
| Boundary | Notes |
|---|---|
| Commercial | Databricks Model Serving for inference |
| GCC | Same |
| GCC-High / IL4 | Azure ML managed endpoints or AKS-MLflow (no Databricks Model Serving in Gov) |
| IL5 (v1.1) | Same as IL4 + Atlas catalog + HSM-CMK |
Federal applicability¶
A federal financial-services agency (Treasury, IRS, USPS, etc.) that monitors transaction streams for fraudulent activity uses this pattern in GCC-High under FedRAMP High + DoD IL4 audit boundary. HIPAA BAA covers if PHI-adjacent (rare for transaction fraud).
Sample Activator rule¶
{
"name": "High fraud-score alert",
"dataSource": {
"type": "adx-kql",
"query": "ScoredTransactions | where ts > ago(15m) | summarize avg(fraud_score) by transaction_id, bin(ts, 1m)",
"splitColumn": "transaction_id",
"cadenceMinutes": 1
},
"rules": [{
"expression": {
"operator": "andStays",
"left": {"operator": "isAbove", "attribute": "avg_fraud_score", "threshold": 0.9},
"durationMinutes": 2
},
"actions": [
{"type": "teams-message", "channel": "#fraud-investigators",
"template": "Transaction {transaction_id} fraud_score {avg_fraud_score} for 2+ min"},
{"type": "databricks-job", "jobId": "fraud-investigation-trigger"}
]
}]
}
Cost (F32 GCC-H baseline for production scale)¶
~$8,500/mo: - Power BI Premium F32: $4,200 - Databricks Premium classic + ML: $2,500 - ADX cluster: $800 - ADLS Gen2: $300 - AOAI (Data Agent + ML feature enrichment): $500 - Misc: $200
Source code¶
examples/fiab-financial-fraud-detection/
Forward migration¶
Standard Loom forward path (OneLake shortcut for Delta tables; dbt + KQL port unchanged). The ML model + MLflow registry exports portably.