Projects

Projects

Anonymised client engagements and method demos. Method demos show how we approach a problem class. They are not named client work.

Showcase

Self-service analytics

Bring your existing backups into a warehouse, then give the team self-service analytics — and an optional AI chatbot — on tools you already have.

  1. Load existing backups into a data warehouse and refresh on the schedule you choose
  2. Plug in a self-service analytics engine and/or a self-service analytics chatbot using AI you already run
  3. Shared dashboards, research, and collaboration from one version of truth
  4. Fixed-scope project that adopts into your stack — written boundary, governed automation, people keep judgment

Ask about a fixed-scope self-service analytics project — info@cbdata.co.za

Anonymised projects

Client engagements and method demos

Client engagement cards are anonymised. Method demos are not named client work.

Client engagement Banking

Enterprise financial data warehouse

The challenge

A financial institution was running month-end reporting across 11 disconnected source systems — core banking, GL, CRM, and loan origination — with a 6-day close cycle, frequent reconciliation failures, and no single trusted view of the balance sheet.

What we built

A centralised cloud data warehouse on Azure Synapse, with automated ELT pipelines ingesting and reconciling data from all 11 source systems nightly. Layered model (raw → staging → conformed → reporting) aligned to BCBS 239 data lineage requirements.

6d → 1dMonth-end close cycle
11Source systems integrated
99.8%Reconciliation accuracy
Azure Synapse dbt Python Azure Data Factory SQL Power BI
Method demo Credit & lending

Credit risk scoring & PD model

Method demo — not a named client engagement. Figures below are illustrative of the method.

Typical problem

Consumer credit scorecards go stale when they sit on old bureau data. Approval rates drift, bad debt rises, and the model no longer matches current borrower behaviour.

Method shown

Rebuild the probability of default (PD) model with gradient boosting on behavioural, bureau, and transactional data. Full model documentation, champion/challenger testing, and a monitoring dashboard for Gini, KS, and PSI.

+18ptsIllustrative Gini lift
-22%Illustrative bad-debt move
<200msScoring latency target
Python scikit-learn XGBoost MLflow SQL Server Power BI
Client engagement Banking

IFRS 9 provisioning & regulatory reporting automation

The challenge

The finance team at a mid-size bank was calculating IFRS 9 expected credit loss (ECL) provisions in Excel each quarter — a 10-day process with version conflicts and audit risk. Regulatory submissions were consistently late.

What we built

An automated ECL calculation engine that stages loans, applies PD/LGD/EAD models per IFRS 9 stage, and produces a provision report with an audit trail. Integrated to the GL for journal generation and connected to the PA reporting template.

10d → 4hQuarterly provisioning cycle
100%Audit trail coverage
0Late regulatory submissions
Python SQL Server dbt Excel (output) Azure DevOps
Client engagement Fintech

Single customer view & 360 data platform

The challenge

A bank with five product lines (home loans, personal loans, vehicle finance, credit cards, and transactional accounts) had no unified view of a customer’s full relationship. Marketing, credit, and collections each used different, conflicting records.

What we built

A master customer data platform with entity-resolution logic to deduplicate and link records across all five product systems. The SCV feeds CRM, the credit risk engine, and a customer profitability view used by relationship managers.

5Product systems unified
+31%Cross-sell conversion rate
SingleSource of truth for all teams
Python Snowflake dbt Apache Spark Tableau
Method demo Credit & lending

Collections intelligence & arrears prioritisation

Method demo — not a named client engagement. Figures below are illustrative of the method.

Typical problem

Collections teams often work a flat call list — every account in arrears treated equally. High-value accounts that are likely to cure get the same attention as low-value chronic defaulters.

Method shown

A collections scoring model that estimates probability of cure within 30 days, weighted by exposure at risk. The model feeds a daily prioritised work queue, with a management view of roll rates, cure rates, and agent load.

+27%Illustrative recovery lift
-35%Illustrative low-value time cut
DailyQueue refresh pattern
Python XGBoost SQL Server Power BI Azure Functions
Method demo Banking

Real-time fraud & AML transaction monitoring

Method demo — not a named client engagement. Figures below are illustrative of the method.

Typical problem

Batch fraud detection on a multi-hour lag means funds have often already moved. Rules-only engines also generate a high volume of false positives that overwhelm compliance teams.

Method shown

A streaming monitoring pipeline that combines a rules engine with an ML anomaly layer. The ML layer learns spending patterns and flags deviations. Alerts route to a case-management view for analyst review.

4h → 1.2sIllustrative detection latency
-61%Illustrative false-positive cut
+40%Illustrative catch-rate lift
Apache Kafka Python Spark Streaming Isolation Forest PostgreSQL Grafana

Client engagement cards are anonymised. Method demos show approach and tooling for a problem class; they are not named client work. Outcome figures on client cards describe that engagement. Figures on method-demo cards are illustrative of the method. No client names, account numbers, or employer systems are listed.

Write to us

If something here is close to your problem, email info@cbdata.co.za. We will take it from there.