Nanzhi Lantern | Auditable Code

Finance manager building finance tools you can audit. Money in Decimal, every number traced to its source. Free calculators:

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

I'm a finance manager who writes code.

I built four free calculators that show their working: tax add/remove, margin vs markup, loan repayment, invoice check.

No sign-up. Runs on your device. Works offline.

https://t.co/qTg3pEVbk3

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Pulled the "latest" image of an open-source tool, then checked the version: v8.8.0-pre, built from the main branch.

"latest" isn't always a release.

For anything holding real records, pin the image to a stable version number.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Question for finance people: what is the smallest rounding difference that ever cost you an afternoon?

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

If I were preparing a finance career for AI:

Month 1: enough SQL to answer one real work question a day.
Month 2: automate one report you rebuild every month.
Month 3: turn one full process into a tool, with check records.

No need to quit to do it.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Work manuals end up everywhere: a Word file, a PDF, someone's chat history. When that person leaves, the know-how leaves too.

I put mine in BookStack, a self-hosted wiki: shelves, books, chapters, pages. Full-text search, every edit versioned.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Merged a folder of Excel files and the total came out too high?

Check each source file for a "Total" row. Drop it after reading, or the merged sum counts every subtotal again.

Then tie it: rows in = rows out, amounts in = amounts out.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

A loan of 50,000 at 9.5% over 36 months: 1,601.65 a month.

The last payment is 1,601.56, because rounding has to land somewhere. Most calculators hide that. Mine prints it.

https://t.co/iWqGymBGVP

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

The only budgeting AI I'd trust runs on my own computer. Bank statements are the last thing I want in the cloud.

Rules first (the subway is travel, no AI needed). A local model only for what the rules can't place. Spot-check 10 lines a month.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

If one of my 7 self-hosted services died, I used to hear it from whoever couldn't open it.

Now Uptime Kuma checks each one every few dozen seconds and alerts me.

The catch: it runs on the same PC it watches. If that PC loses power, nobody gets an alert.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Before trusting SQL an AI wrote, run it on a period you already know the answer for.

If last month's total doesn't match the signed-off report, the query is wrong, however clean it looks.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Seven rules I follow when AI helps me build finance tools:

Money is Decimal. Raw exports are read-only. Every number traces to its source. Tests before trust. A person approves anything that moves money. Secrets stay out of code.

AI drafts; I audit.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Finance tasks AI is shrinking: rule-based entries, ticking reconciliations, formatting the same monthly report.

Tasks worth more now: deciding whether to pay, explaining why profit moved, checking whether the AI's numbers are right.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Fixed asset register in Excel: who has which laptop is known only by whoever keeps the sheet.

Snipe-IT tracks tag, serial, cost, check-out and depreciation.

The step most guides skip: export it monthly and tie total cost and accumulated depreciation to the ledger.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Python's round(2.5) returns 2, not 3.

It rounds exact halves to the nearest even number. Fine for statistics, wrong for most invoices.

For money: Decimal, ROUND_HALF_UP, and say so in the code.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Invoices get re-added by hand more often than anyone admits.

I made a checker that does it line by line, then compares with the total printed on the invoice and shows the difference.

https://t.co/MFWI58paRB

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Online PDF tools want the file uploaded first. Invoices, contracts, bank statements: I'd rather not.

Stirling-PDF runs in the browser on my own PC. Merge, split, OCR, compress.

The catch: it's the memory hog of my setup. 896MB, against 24MB for my password vault.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Three things I do the day I install any self-hosted tool:

1. Change the default admin password.
2. Turn off public sign-ups.
3. Never open the port to the internet.

Several tools I tested ship with at least one of these wide open.

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

What AI took off my desk this year: ticking reconciliations, merging files, first drafts.

What it didn't: deciding whether to pay, explaining a variance, signing the report.

What is the most repetitive hour in your week?

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Same $20 profit, two different answers:

Cost 80, price 100.
Margin = 20 / 100 = 20%
Markup = 20 / 80 = 25%

Mix them up in a pricing sheet and every product is off.

Calculator: https://t.co/J3nD4IV4Zi

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

Borrow 10,000. Repay 1,000 a month for 12 months.

"2,000 interest on 10,000, so 20%." Not quite. You repay principal every month, so you never hold the full 10,000.

Real rate: 2.92% a month, 35.07% a year, 41.30% compounded.
Excel agrees: =RATE(12,-1000,10000)*12

[Nanzhi Lantern | Auditable Code @Edgar_cui1]

The SQL bug I see most in finance queries:

Join invoice headers to invoice lines, then SUM the header total. An invoice with 5 lines gets counted 5 times.

10-second check before any SUM:
COUNT(*) vs COUNT(DISTINCT invoice_id)
If they differ, your rows have multiplied.