Projects Built with AI

Tools Matthew has built himself, with AI as the second pair of hands — not available anywhere else but his own GitHub. Each write-up covers why and how it was built, not just what it does; each repo carries the usual documentation set — GUIDE.md (getting started), MANUAL.md (feature reference), DEVELOPMENT.md (architecture and how to build/contribute) — linked from its page. Where a project has a working deploy, it’s linked here too. For AI tools and products he uses but didn’t build, see External AI Tools.

Money & Admin

Split Ledger

Works out who pays whom after a shared expense, in the fewest possible payments — and can say why no shorter set of transfers exists. The algorithm is formally proved minimal in Lean 4; the page is a transcription of it that re-checks every answer against the same condition the proof is about.

Key Features:

  • The fewest transfers, not just a correct settlement
  • Whole-number shares, so “two thirds and one third” is exact
  • Money in integer pence — no floating-point drift
  • A certificate check on every answer, with a badge saying what it means
  • Share a ledger by link; nothing is stored on any server

Best For: Holidays, house shares, group presents, estates — where people paid different amounts

Use it here →


Astrology

Astrodynamics

A local-first astrology app: offline chart calculation, interactive SVG wheels, Placidus houses, transits, synastry, and midpoint composites. Installable as a PWA; nothing leaves the device.

Key Features:

  • Offline ephemeris checked against Swiss Ephemeris and JPL Horizons
  • Full chart maths: houses, aspects, transits, synastry, midpoint composites
  • Four switchable interpretation voices with a plain-language glossary
  • Installable PWA, fully offline after the first load

Best For: Personal chart work, transits and synastry, offline astrology reference

Use it here →


Finance

ledgerforge

Turns a folder of bank statement exports (OFX/QIF/CSV) into a self-checking double-entry GnuCash book — multi-currency, coded chart of accounts, and a residual-sum check that tells you exactly when something’s wrong.

Key Features:

  • One parser per bank format; a rules engine that categorises every transaction
  • Transfer detection, so moves between your own accounts never count as income or expense
  • Self-checking: the signed total across every account is exactly nought when the book is right

Best For: Turning bank exports into a trustworthy, auditable account book

Use it here →


Reading Tools

adaptive-text

Lets a reader dial a piece of writing between summary and full detail, and adjust its formality and reading level, instead of reading one fixed version.

Key Features:

  • A resolution slider spanning summary to full detail on the same source text
  • Adjustable formality and reading level
  • Mock mode (no API key) and a real, OpenAI-backed mode

Best For: Exploring whether one source text can serve very different readers


Language

miolingo

A multi-language pronunciation trainer that scores learner speech phone-by-phone against articulatory feature distances, not naive text-edit distance.

Key Features:

  • Phone-by-phone scoring via real articulatory feature distances
  • Real ASR in the loop (Whisper) compared against espeak-ng’s reference IPA
  • Per-language accent tolerance derived from espeak-ng’s own allophone rules

Best For: Serious pronunciation practice across seven languages


Maths

mathematica-in-lean

A Wolfram bridge for Lean 4: Wolfram proposes a certificate, Lean’s kernel independently checks it — so no proof trusts Wolfram’s output directly.

Key Features:

  • Wolfram-assisted certificate discovery, kernel-checked in Lean
  • Ported to Lean 4 / mathlib4
  • Flagship result (∑ C(n,k)² = C(2n,n)) is sorry-free and axiom-clean

Best For: Seeing what CAS-assisted, kernel-verified mathematics looks like in practice


About This Directory

Every project here is Matthew’s own — built with AI, but the judgment, the review, and the merges are his. Each write-up is a companion to the deeper documentation in the project’s own repo, not a replacement for it.