Every tax season, bookkeepers and small business owners hit the same wall: banks lock 12–36 months of historical statements inside PDFs, while existing PDF-to-CSV tools either charge $29/mo+ recurring subscriptions for an annual task or ask you to upload raw financial PII to remote cloud servers.
Meanwhile, throwing bank statements at an LLM creates a worse problem: probabilistic math. Accounting has an exact ground truth — books either balance to the penny or they don't.
Here is the architecture behind Local Bank Statement Reconciler (https://vz-001-reconciler.pages.dev): a pure client-side TypeScript + Vite web app that parses bank PDFs, verifies ledger math deterministically in integer cents, and exports CSV, Excel, and QuickBooks .qbo (OFX) with zero backend servers and $0 hosting cost.
1. Architecture Overview: Why Pure Client-Side Wins on Trust and Cost
| Layer | Choice | Why |
|---|---|---|
| PDF Parsing | pdfjs-dist (bundled locally) | Zero external CDN worker requests; raw PDF bytes never leave browser RAM |
| Reconciliation | Deterministic integer-cents engine | Verifies Opening + Credits - Debits == Closing with zero IEEE-754 float drift |
| Exports | Pure JS RFC-4180 CSV, OFX .qbo, SpreadsheetML .xml | Zero server dependencies; works offline |
| Licensing | WebCrypto ECDSA P-256 (VZ1.<payload>.<sig>) | Offline asymmetric signature verification — zero auth server needed |
| Hosting & Checkout | Cloudflare Pages Free + itch.io | $0/mo static hosting with zero-domain external checkout |
Even the built-in "Try Built-In Sample Statement (<10s)" button generates a real PDF 1.4 byte stream in memory using TypeScript (new Uint8Array(...) -> File), so a visitor watching F12 -> Network sees literally 0 HTTP requests from click to CSV export.
2. Coordinate Clustering & Multi-Line Row Reconstruction
PDFs don't store tables — they store positioned text runs (str, x, y). Bank statements complicate this in two ways:
- Numeric columns may jitter by
1–3 ptvertically relative to the date/description. - Long merchant descriptions wrap onto 2–3 continuation lines below the transaction date.
We normalize every text run through pdfjsLib.Util.transform(viewport.transform, item.transform), cluster rows within an adaptive yTol = clamp(lineHeight * 0.45, 2.0, 5.5), and merge intra-row glyph spans into column tokens. Once header bands (Date, Description, Debit, Credit, Amount, Balance) establish X-coordinate intervals, any subsequent row that lacks a date and numeric amount is stitched onto the preceding transaction's description rather than dropped or emitted as an orphan row.
3. The 0-Glyph Scanned PDF Guard & Integer-Cents Invariant
Two silent failure modes destroy user trust in PDF converters:
- Passing an image-only scanned PDF through a text parser -> emitting
0rows -> computing0 + 0 - 0 == 0-> falsely reporting"Reconciled!". - Accumulating floating-point dollars (
0.1 + 0.2) and missing penny discrepancies.
Our parser counts selectable non-whitespace glyphs (totalGlyphs) and inspects page.getOperatorList() for image XObjects (paintImageXObject). If totalGlyphs === 0, the UI halts with an actionable warning explaining why we refuse to upload scanned images to a cloud OCR server. All transaction sums run strictly in integer cents (19.99 -> 1999).
4. Offline Asymmetric Licensing via WebCrypto ECDSA P-256
To keep the $19 Pro upgrade 100% offline-capable (no license check-in server, no tracking), the app verifies signed VZ1.<payload_b64u>.<sig_b64u> tokens locally using window.crypto.subtle and an embedded ECDSA P-256 public JWK (x and y coordinates only — no private scalar d ever ships in client code).
Try It & Inspect the Network Tab
- Live Tool:
https://vz-001-reconciler.pages.dev - Test in < 10 Seconds: Open
F12 -> Network, click "Try Built-In Sample Statement", and verify0network requests from parse to CSV export. - Pricing: Core single-statement parsing, full balance check, and up to
30exported CSV rows are$0forever (no login). A$19one-time Pro license unlocks unlimited rows, multi-month PDF merging with duplicate/gap detection, Schedule-C categorization, QuickBooks.qbo(OFX), Xero CSV, and a multi-sheet Excel audit workbook.
