1. Executive Summary & Problem Statement:-Financial institutions face severe operational bottlenecks when detecting coordinated money mule rings and rapid-drain fraud schemes. Fraud syndicates exploit real-time payment rails (such as UPI and IMPS) by siphoning illicit capital across networks of newly onboarded, colluding accounts within minutes. Traditional fraud surveillance relies on static batch pipelines and fragmented relational stores. Fraud operations teams are forced to manually construct complex multi-table SQL queries across account databases, transaction streams, and device logs. This operational lag creates multi-hour delays in freezing illicit fund flows, resulting in direct financial loss and missed regulatory compliance deadlines. AegisBank Copilot provides an autonomous, Lakehouse-native fraud surveillance solution powered by Databricks Unity Catalog, Databricks Genie, and Databricks Apps (Streamlit).
2. Solution Video Demo & Artifacts:-
3. End-to-End System Architecture:-
[Raw Ingestion Streams & Telemetry]
โ
โผ
[Bronze Layer: Append-Only Delta Tables]
โโโ raw_accounts
โโโ raw_transactions
โโโ raw_device_fingerprints
โ
โผ
[Silver Layer: Cleansed & Deduplicated]
โโโ clean_accounts (SCD-type / Current State)
โโโ clean_transactions (Timestamp enrichment & validation)
โโโ clean_device_fingerprints (Normalized device hashes & IP subnets)
โ
โผ
[Gold Layer: Business Intelligence & Risk Rules]
โโโ dim_accounts
โโโ fact_transactions
โโโ fact_device_fingerprints
โโโ fact_fraud_signals_gold (Mule Risk Scoring & Velocity Flags)
โ
โผ
[Genie Agent & Databricks App Frontend]
โโโ Unity Catalog Semantic Metadata & System Instructions
โโโ Conversational Natural Language to Spark SQL Generation
โโโ Compliance Dossier (SAR) Auto-Generation & Debit Freeze Actions
4. Lakehouse Data Foundation (Unity Catalog):-
The surveillance pipeline organizes and governs 4 core Gold tables inside the catalog aegis_fraud_workspace and schema fraud_surveillance_lake:
dim_accounts: Master account registry capturing account IDs, onboarding channels, and KYC verification statuses.
fact_transactions: Granular financial transactions across UPI, IMPS, NEFT, and Card payment rails with real-time velocity indicators.
fact_device_fingerprints: Device identifiers, IP subnet telemetry, and browser fingerprints used to pinpoint hardware-level collusions.
fact_fraud_signals_gold: Enriched feature store evaluating rapid-drain dynamics, composite mule risk scores (0โ100), and collusion risk tiers (HIGH, MEDIUM, LOW).
5. Conversational Genie Copilot Integration:-
By embedding domain-specific semantic instructions and multi-table entity relationships inside Databricks Genie, investigators can query the Lakehouse conversationally without writing manual joins:
Mule Ring Detection: Rapidly aggregates accounts linked to shared hardware signatures (device_fingerprint).
Rapid-Drain Surveillance: Instantly queries newly opened accounts (<30 days old) receiving high-value inflows exceeding threshold limits.
Transparent SQL Lineage: Every natural language question outputs the exact, deterministic Spark SQL join used to produce the result, ensuring complete compliance transparency.
Benchmark Genie Prompt Example:
"Show all accounts created in the last 30 days that received transactions exceeding 10000 on UPI".
6. Databricks App UI & Compliance Operations:-
Deployed directly on Databricks Apps, the copilot frontend provides fraud teams with a centralized command center:
Executive Metrics Dashboard: Real-time visibility into active mule rings, flagged capital volumes, and average resolution latency.
Investigation Action Chips: Pre-built surveillance workflows for single-click anomaly drill-downs.
SAR Dossier Auto-Generation: Instant generation of formal Suspicious Activity Report compliance summaries for the Financial Intelligence Unit (FIU).
Active Remediation: Simulated automated account debit freeze execution via core banking webhooks to immediately stop fund siphoning.
---
**Tags:** #Databricks #DatabricksGenie #DatabricksApps #UnityCatalog #DeltaLake #FraudAnalytics #FinTech #TrackA
