Start with the need.
Financial exports arrive in different formats and can contain missing fields, duplicate candidates and inconsistent references. Analysis is more useful when those defects are visible before anyone relies on the totals.
My contribution
I built a Python and SQL pipeline that regenerates synthetic records, standardises them into SQLite, runs data-quality checks and produces exception and KPI outputs. I added reconciliation to independent control totals and documented a downstream Power BI workflow.
How it works
Conceptual flow. Implementation details and reproducibility instructions are in the repository.
A closer look at the work.
| Check | Recorded result | Next step |
|---|---|---|
| Critical-field completeness | 130 transactions | Review missing amounts, dates or vendor IDs |
| Vendor references | 79 transactions | Investigate references absent from the vendor master |
| Duplicate screening | 290 candidates | Corroborate with transaction-level evidence |
These are saved results from the synthetic project. Candidate matches do not establish confirmed duplicate payments.
Inspect the full findingsThe decisions behind the code.
Make defects reproducible
Fixed-seed synthetic data includes deliberately injected issues so the checks can be demonstrated repeatedly.
Separate flags from conclusions
Duplicate and Benford screening prioritise investigation; neither establishes fraud or a confirmed duplicate payment.
Reconcile before reporting
Ledger totals are compared with independent controls using a documented £0.01 demonstration tolerance.
What the evidence says.
The repository documents 10,150 synthetic transactions, completeness and reference checks, a £0.01 reconciliation tolerance, and CSV outputs for exceptions, KPIs and reporting. Findings are demonstrations on synthetic records.
Public repository README, reviewed 17 November 2026.
Read the source documentationWhere the boundaries are.
- No client financial data is used and no recovered money or business impact is claimed.
- Screening signals need transaction-level review and business context.
- The reconciliation tolerance is a demonstration choice; a real engagement would require an agreed control policy.
What I learned
A useful data pipeline does more than transform rows. It makes the checks, exceptions and next investigation steps understandable.