Case studies / Data foundations

Data foundations · Credit-decisioning platform provider

A secure, multi-tenant credit-decisioning data lake, built from heterogeneous source files

Credit-decisioning data arrived as heterogeneous files across multiple tenants, with no canonical model and personal data mixed into operational data. We built one queryable model — with isolation by design.

5
Data layers

Secure landing, PII-stripped raw, staging, canonical and reporting layers.

7
Domain entities

Applications, applicants, scores, adverse actions, collateral, financials and business entities.

Sector
Credit-decisioning platform provider
Engagement
Multi-tenant data lake build
Problem area
Data foundations
Confidentiality
Details anonymised

The challenge

Heterogeneous files, mixed-sensitivity data, no shared model

Credit-decisioning inputs arrived as heterogeneous files from multiple tenants, with no canonical data model, and personal data commingled with operational data — a real risk for a financial-services platform.

Before

  • Heterogeneous source files, tenant by tenant
  • No canonical, queryable model
  • Personal data mixed with operational data
  • Manual handling slowed decisioning

After

  • One canonical model, queryable in SQL
  • Personal data isolated by design
  • Multi-tenant architecture built for scale
  • Deployed via IaC across Dev, UAT and Prod, with the client's engineers trained to run and extend it
Delivery status

Foundation delivered and validated against the client's existing reporting. Client-side data onboarding is currently paused while they sequence bringing their own data onto the platform — that's adoption sequencing, not a delivery gap. No usage claims until it's switched on.

How we approached it

Isolate first, then unify

01

Model the domain, not the files

We made one canonical credit-decisioning model the fixed target that every client's source files map into.

02

Make PII a design constraint

Layered isolation, hashing and “no PII queryable in the lake” formed a hard acceptance criterion, not an afterthought.

03

Onboard by metadata, not code

Lookup-table-driven mappings and a shared flattening job were deployed via IaC, with client-team training included.

Under the hood

AWS S3AWS Glue (Python)AWS BatchAWS LambdaStep FunctionsAmazon AthenaCloudFormationAmazon SageMakerApache IcebergSnowflakePower BI

Next step

Handling sensitive data across multiple tenants?

We build isolation into the foundation itself — so it doesn't depend on policy alone.

Start a conversation →