Maria Alftaih

Personal Finance Ledger

A self-hosted tool for understanding what you actually spend, built for Turkish credit card statements and Turkish inflation. It imports statements from PDF, works out which lines are the same purchase printed twice, sorts spending into categories you control, and reports in real terms.

• Reading a table out of a PDF — PDFBox flattens columns into an indistinguishable word stream, so the importer keeps x-coordinates and derives each column from where its header label starts. Two bank layouts (Garanti BBVA, Yapı Kredi) that differ in every way that matters, plus a CSV fallback.
• Money is a type, not a double — a BigDecimal and a Currency, scale fixed at 2, throwing on cross-currency arithmetic instead of coercing. It crosses the JSON boundary as a string, so the browser cannot quietly turn it back into a float.
• Turkish is a hostile locale — i↔İ and ı↔I mean "TITLE".toLowerCase() returns tıtle on a tr-TR machine and every string match silently stops working. Matching folds with Locale.ROOT, display with tr-TR, and a test asserts the trap is real.
• Deduplication with a hard constraint — consecutive statements reprint the days they share, but two identical purchases on the same day are legitimate, so a unique key would erase real spending. jqwik property tests pin the invariants: deduplicating a statement against itself is a no-op, and merge order does not matter.
• Real terms, not the headline rate — reports divide CPI index levels rather than adding the annual percentage, which is the easy mistake and answers a different question. The index lives in a local table, so producing a report never makes a network call.
• A boundary the build enforces — Maven Enforcer bans Spring, JPA and Hibernate from the domain and import modules transitively, checked in CI on every push; integration tests run against a real PostgreSQL image under Testcontainers rather than H2, and CI fails if they were skipped.