A governed lakehouse for a bank, layer by layer
How data moves from a bank’s source systems to reports, models and agents on Databricks, and which control records access at each step.
Six layers, with governance alongside each one
Each box names the job and the Databricks product that does it. Source system names are generic labels.
Sources
- Accounts and transactionsCore systems
- CardsCard platform
- Loan applicationsLoan origination
- Customer recordsCRM
- Market dataMarket feeds
Ingestion
- Built-in connectorsLakeflow Connect
- FilesAuto Loader, UC volumes
- Event streamsStructured Streaming
- Database changesAUTO CDC
- Query in placeLakehouse Federation
Bronze
- Raw copy with load date and batch IDDelta tables in Unity Catalog
Silver
- Cleaned, with change history (SCD-2)Lakeflow Spark Declarative Pipelines
- Quality checksExpectations
Gold
- Risk and credit
- Customer 360
- Regulatory reports
Serving
- Reports and dashboardsDatabricks SQL, AI/BI
- Questions in plain languageGenie
- ModelsMLflow, Model Serving
- AgentsAgent Bricks
- Operational appsLakebase
Governance
- Grants, row filters, column masksUnity Catalog
- Sensitive data tagsGoverned Tags
- Lineagesystem.access.table_lineage
- Access logsystem.access.audit
- Quality monitoringData Quality Monitoring
- Model and agent trafficAI Gateway
- Cost by teamsystem.billing.usage
What Vietnamese law asks, and where the architecture answers
Card data, transaction history and credit information are sensitive personal data, so most core tables on a bank’s platform fall under the strictest rules.
| Requirement | Legal basis | Where it is handled |
|---|---|---|
| Card details, account transaction history and credit information are sensitive personal data | Decree 356/2025, Article 4(k) | Governed Tags mark sensitive columns; column masks and row filters in Unity Catalog |
| Multi-factor authentication, access rights, monitoring for abnormal access | Law 91/2025, Article 30(3); Decree 356, Article 9 | Group-based grants; system.access.audit |
| Encryption at rest and in transit for personal data in the cloud | Decree 356, Article 12 | Security settings of the workspace and of the bank’s own cloud account |
| Carry out a deletion request within 20 days | Decree 356, Article 5 | Lineage finds every table holding the person’s data; DELETE, then VACUUM on Delta tables |
| A workspace outside Vietnam counts as cross-border transfer | Law 91/2025, Article 20(1)(c) | Region chosen at the start; the impact assessment draws on lineage and audit records |
| An authorised person reviews automated credit decisions and high-value transactions | Decision 33/2026; Decree 142/2026, Article 8(2)(b) | Approval step in the agent, MLflow tracing, system.access.audit |
Legal bases checked against the source texts in our note on the Personal Data Protection Law and our note on Decision 33. This page does not replace legal advice.
Four use cases on the same foundation
Each use case adds its own Gold tables and models, and reuses the ingestion and controls already in place.
Fraud detection
Card and transaction events arrive as streams; models are tracked in MLflow; investigators see only the columns their role allows.
AML, KYC and regulatory reporting
Reports come from Gold tables with lineage back to the source systems, and every read is written to the audit log.
Credit risk
Scoring models carry lineage from features to source data, and the credit officer’s approval is recorded.
Customer 360
Customer records are merged in Silver; business teams ask questions through Genie within the access they have been granted.
Sample problems by sector, and the package that fits
Typical problems in banking and adjacent sectors, each matched to the package whose scope suits it.
| Sector | Problem | Package |
|---|---|---|
| Banking | Transaction monitoring for anti-money laundering (AML): alert rules run on governed transaction data, and every alert traces back to the transactions behind it. | Governed Pilot |
| Banking | Shared financial metrics such as net interest margin (NIM), cost-to-income ratio (CIR) and non-performing loans (NPL), defined once and used by both reports and Genie. | Analytics Use Case Sprint |
| Insurance | IFRS 17 reporting and claims analysis on one data model, with lineage from policy and claim records to the figures in each report. | Analytics Use Case Sprint or AI & Agents to Production |
| Energy group | Consolidated reporting across member companies: each company’s ledger lands in one governed model, and intercompany eliminations are recorded and traceable. | Governed Pilot |
Scope and price are set out on each package page. The right package is confirmed after the discovery session.
Compare your platform with this architecture
A readiness review scores your Databricks workspaces on governance, lineage, reproducibility and cost. The free 90-minute discovery session is the first step.
Talk to us
Tell us what you are working on. A named person replies within one business day.