Econometrics student in Antalya, Türkiye, working as a court clerk at the Ministry of Justice. I write Python tools for problems where being confidently wrong is worse than returning nothing.

Both projects below are built on one idea: don't warn about the wrong answer, make it unreachable.

Detects and masks personal data in Turkish text, so a document can be processed, or sent to a cloud model, without its identifiers leaving the building.

Check-digit validation for the identifiers that have one. A locally run model for names and addresses, which no pattern can find. Masked output is re-scanned, and a surviving identifier raises an exception rather than ship a leak quietly.

Measured over 220 generated documents: 100% on every pattern-matched type, 99.5% on names, 0 false positives on 30 control sentences built around near-miss traps.

Every one of those numbers started out worse. Making the test harder is what found the real problems: names written without a label, addresses that were being masked only halfway, and a model configuration that silently returned nothing at all.

An MCP server over the Central Bank of Türkiye's statistical database, so a model can pull Turkish macroeconomic data and analyse it without inventing series codes.

It has no raw correlation tool, deliberately. USD/TRY against CPI correlates at 0.99 in levels and 0.57 log-differenced at a one-month lag. The first figure is spurious, the lag is not contemporaneous, and a tool that hands over the first number without saying so is worse than no tool.

Open to remote backend and data engineering work.