← All work

Automation · data pipelines · reconciliation

End-to-end personal finance automation

Statements, emailed receipts and brokerage exports arrive in different formats, at different times, from different sources. This ingests all of them into one store and reconciles them through a tiered matcher — with one rule that decides everything: a receipt is evidence, never a financial event.

Everything below is measured from the running system — 2,214 records spanning May 2025 to November 2026. Counts and structure only: the amounts, the vendors and the account numbers are the part that stays private, which is a constraint the design had to respect from the start rather than one added at the end.

The problem

Recording a transaction is easy. Knowing that this evidence belongs to that charge is the whole problem — and it is not a parsing problem, it is a judgement problem wearing a parsing problem's clothes.

Three things arrive independently and none of them agree. A receipt is emailed on the day of purchase and quotes a pre-tax figure. The card charge lands days later, includes the tax, and names a parent company rather than the shop. A brokerage export arrives at month end, in a different currency, with its own idea of what a date is. And in every set there is one payment that must never be allowed to settle two bills, because the fastest way to make a ledger wrong is to make it look tidy.

Reconciling that by hand is the kind of work that quietly never gets done.

How it fits together

Four sources, one store, and a matcher in the middle that is doing the actual thinking.

The ingestion path: four sources into one store, then reconciliation, then a dashboard Emailed receipts, bank CSV exports, brokerage activity exports and manual entries all normalise into a single SQLite store. A tiered matcher then reconciles evidence against charges, and a self-hosted dashboard reads the result. Sources — they disagree by design emailed receipts pre-tax, day of purchase bank exports post-tax, days later brokerage activity month-end, multi-currency manual entries the fallback that stays One normalised store SQLite · every source parsed on its own terms tiered matcher → anomaly detection → self-hosted dashboard then a person looks
Nothing is merged at ingest time. Sources are parsed on their own terms and kept distinguishable, so the matcher can always tell what kind of evidence it is holding.

The matcher: six tiers, ordered by how much they assume

Each tier is a different theory of how two records can be the same thing. They run in order, and the order is the design: the cheap, safe theories first, so a confident match never gets made by a fuzzy rule that happened to fire first.

The six matching tiers, each reached when the previous one cannot decide Exact match first. If the amounts differ by the tax rate, a tax-adjusted match. If the currencies differ, a foreign-currency match at the day's rate. Then fuzzy matching, then batch matching for grouped charges, then split matching for one payment covering several items. A final invariant check rejects any result where one payment would settle two bills. 1 · exact same amount, same day 2 · tax-adjusted receipt said X, charge says X + tax 3 · foreign currency matches only at that day's rate 4 · fuzzy the vendor writes its own name three ways 5 · batch one charge, several items behind it 6 · split one payment, several bills cannot decide — next tier The invariant that outranks all six A receipt is evidence, never an event. A charge is the event. The receipt only tells you what the charge was for. So the matcher is allowed to fail. What it is never allowed to do is let one payment settle two bills — a wrong ledger that looks tidy is worse than an obvious gap. Anything unmatched surfaces for a person, rather than being guessed at.
The correctness rules mattered more than the parsing. Every tier exists because a real pair of records broke the tier above it, and the last line of defence is an invariant check that rejects a plausible match rather than accepting an impossible one.

What it holds, measured

Two charts, both counts rather than amounts. The first is the volume of records the store has accumulated; the second is where they came from — which is really a chart about how many different formats one system had to learn.

Records ingested per month, May 2025 to November 2026 Monthly record counts rise from around sixty to seventy in mid-2025 to a peak of 222 in September 2026. January 2026 shows a step change from 75 to 103 as new sources were added. The final bar, November 2026, is partial at 11 records. Records ingested per month peak 222 · Sep 2026 May 2025 Jan 2026 Sep 2026 Nov a step change — sources came online together faded bar = month in progress
2,214 records. The growth is in what the system learned to read, not in spending — and the November bar is drawn faded because the month is not finished, which is the sort of thing a chart should say rather than hide.
Ingested records by source format Bank CSV exports account for 1,330 records, brokerage activity exports 431, brokerage holdings 267, and email-sourced records 186. Records by originating format — four parsers Bank CSV 1,330 Brokerage activity 431 Brokerage holdings 267 Email / other 186
Four formats, none of them introduced by choice. Each one arrived because the alternative was re-keying it.

What made it interesting

Not the parsing. Reconciliation is where the edge cases live, and every one of them is a case where the correct answer is less convenient than the plausible one:

Building it changed how I think about automation generally. The interesting part is never the happy path; it is deciding what the system does when it cannot be sure — and the answer here is that it says so.

Outcome.

The dashboard

The pipeline's output still has to be readable by a person, so it ends in an interface. One self-contained HTML file — no build step, no framework, self-hosted fonts — reading the same SQLite store, installable to a phone home screen, and organised as three views: Dashboard, Ledger, Receipts.

The dashboard opens with a snapshot, and then two cards that are the actual point: Wins and Opportunities. Not more figures — the reading of them. Below that, four charts (spending by category across twelve months, this month, the savings-rate trend and its projection, the net-worth projection), then bills, a watchlist, a portfolio, and a drop zone for importing statements.

The finance dashboard: a snapshot row, Wins and Opportunities cards, and a spending-by-category chart. Every dollar figure is blurred, and the charts with them.
Values blurred — with the dashboard's own control. It is the only honest way to screenshot it, and the feature is the more interesting story. It is dark by design: its ground is #0b0910, the same night value the rest of this site uses.

The blur button, and why it is the part worth keeping

There is a button that hides every dollar figure, and it exists for a reason that is not obvious: so the dashboard can be shown to someone. Every money value is wrapped at runtime and blurred; the charts blur with it.

Two decisions inside that small feature are better than the feature itself:

That is a small piece of engineering, and it is the difference between a demo and something a person keeps using. Its existence tells you the author expected to want to show it to somebody — which is a design assumption you can build a feature on, and the reason this page can show you the interface at all.

Facts, in a form you can check

StoreSQLite, single file, 11 tables
Records2,214, spanning May 2025 to November 2026
SourcesFour formats: bank CSV, brokerage activity, brokerage holdings, email
MatchingSix ordered tiers, cheapest and safest first
IngestionEmail receipts parsed directly; every source de-duplicated on the way in
DiagnosticsAnomaly detection over the reconciled result, run on a schedule
InterfaceOne self-contained HTML file, no build step; a dashboard, a ledger and a receipts view; installable to a phone home screen
SharingA blur control that hides every money figure on demand — runtime-tagged, and deliberately not persisted
Data policyCounts and structure only. No amounts, no vendors, no account numbers — in the system's public description and in this page alike