Skip to content

← Selected work

Counting House

An open-source, local-first personal finance tracker for the UK (50+ stars on GitHub). Statements, CSV exports and screenshots go in; git-versioned JSON comes out: estate value, spending, tax year and projections.

Role
My own open-source project
When
09/2026 – Present
Stack
TypeScript, React, Hono, Go
Licence
MIT

The extraction eval

I built an eval that measures, field by field, how well the import pipeline reads synthetic documents: CSV exports, PDF statements and phone screenshots.

  • Claude Sonnet scored 100% on 28 cases for $1.04.
  • On a later, 41-case version, local models served through my own llama.cpp gateway scored 99.1% at no API cost.

Reading the two scores

They come from different runs on different versions of the eval, so they are not a head-to-head comparison.