I find the place where two sources stop agreeing.

A CEO dashboard that matches the official statements

14K+sales and 33K bank transactions, 2020 to 2026

EFT GROUP, agricultural infrastructure distributor. Remote, current work. 1C:Enterprise, Python, Polars, CatBoost, LightGBM, Chart.js.

What I built

A dashboard the CEO uses daily: profit and loss, cash flow, products, clients and stress scenarios. It is fed by extractors I wrote against the company's accounting, sales and orders registers, with validation on every load.

What did not agree

The dashboard showed a gross loss for 2025. The statutory financial statements showed a profit.

How I found it

The sales register already stores revenue net of VAT. The extraction subtracted VAT a second time, so every sale looked smaller than it was.

What changed

After fixing the query and rebuilding the data, revenue and gross profit came back in line with the official statements, and the gross profit forecast error fell from over 100% to about 20%.

The dashboard is private. Screenshots on anonymized data are available on request.

Contracts credited to the right supplier

863contracts credited to the wrong company, found and fixed

AusTender data platform, public study project. Snowflake, dbt Core, SQL, Power BI, GitHub Actions.

What I built

A data model joining 241,000 Australian federal contracts to a 20.4-million-entity business register, with automated tests on every change and a published data dictionary and lineage graph.

What did not agree

Every test passed, yet one business number carried 140 unrelated supplier names, and 863 contracts were credited to a company that never won them.

How I found it

I stopped trusting the obvious identifier and checked it against the national business register as a second source, which separated real suppliers from an agency number used as a placeholder.

What changed

The model was re-keyed, the evidence documented, and a test added so the same mistake cannot return silently.

Questions about payments, answered exactly

40Ktransactions extracted from statements and documents

Internal tool over company bank statements. PostgreSQL, Python, Docker.

What I built

A tool that answers questions such as "how much did we pay this supplier in March" over bank statements in two export formats.

What did not agree

A language model asked to add up payments quietly undercounted whenever there were more records than it could read at once.

How I found it

Numeric questions now go to an exact SQL aggregation instead of the model. Counterparties are matched by tax ID first and by fuzzy name matching only as a fallback.

What changed

Totals match hand-checked figures across both statement formats, and every release must pass a fixed set of known questions before it ships.

The tool runs on company data and is private.

ETF portfolios tested on unseen data

4portfolio strategies tested on unseen data

Mean-variance analysis in Python comparing four strategies out of sample, with return, volatility, maximum drawdown and value at risk.