The problem
The tools people reach for to sort out a bank statement, polish a draft or answer questions about a document
increasingly mean uploading the file to someone's server, and an answer that can change from one try to the
next. For a bank statement, a draft or a family photo, many people would rather not do either. A lot of this
work does not need a prediction at all: it needs careful measurement, and a check that the result is right.
What it does
- Bank statements: reads the PDF your bank gives you (or a CSV), finds merchants charged many times,
subscriptions billed twice, money in and out of advance and pay-later apps, payments returned unpaid and
fees, and keeps moves between your own accounts apart. It then checks what it read against the
statement's own printed totals, to the cent, and says so if they ever differ.
- Humanizer: takes text an AI wrote and makes it read like a person wrote it: plain words for stock
phrasing, empty lead-ins removed, contractions used, long sentences split. Every number, name and
quotation is counted before and after, and a rewrite that would change one is held back.
- Pictures and sound: removes what rides along in a file (where a photo was taken, the software,
comments), then decodes both versions and confirms every pixel, or every sample, is identical.
- And more: questions answered from your own documents with the source line (or "your documents do
not say"), coverage of one event split into sides with what all sides agree on, a code checker that proves
runaway patterns by timing them, and text fixes and word-for-word summaries.
Who did what
- Me: the concept and every decision: the design rules (input, transform, output; each part knowing only
its neighbours; the work built in a fixed sequence of steps), which tools to build and for whom, what each
one may and may not change, and the call to show results people can check rather than claim them.
- The AI (Claude, in VS Code): the code, the tests and the site, under written standards, with a plain
report of what it checked and what it did not.
How the work was kept honest
- Tests that try to break it: 95 hostile edge cases (empty input, huge files, emoji, damaged and
wrong-type files, malformed statements) and 6,000 randomized texts in which no number, name or quotation was
ever changed.
- Testing found real bugs, and each is now guarded by a test: one step took 90 seconds on a large
document and now takes a tenth of a second; four inputs that crashed now get plain messages.
- Proved, not guessed: the code checker times each risky pattern on longer and longer input; on patterns
shaped like real published advisories it flags the three that truly run away and none of the four
look-alikes, and it raised no false alarm on 606 files from widely used packages.
- Real data: the statement reader was tested on a real bank statement and matched its totals exactly,
without the statement leaving the computer or entering any repository.
- Checkable privacy: a recorded network audit shows the page sends nothing while the tools run; the site
invites visitors to watch their browser's network panel and see the same.
Why it matters to an employer
- AI-assisted development done properly: fast output from the AI, with direction, standards and checks
from a person.
- Product judgment: tested with real data, which exposed that the first version could not read the files
banks actually hand out; that changed the design.
- Evidence over assertion: results that verify themselves, honest limits, and no claims a visitor cannot
check.
- Privacy by design: no accounts, no uploads, no stored data; the work happens on the visitor's device.
Technologies and skills
- JavaScript (ES modules) shared between a desktop app and the website, packed for the browser with
esbuild; the same tested code runs in both.
- pdf.js for reading statements by position on the page; Web Workers for timing patterns
safely; an abstract syntax tree parser (acorn) for reading code.
- WebGL: the hero image is the Schwarz Diamond surface drawn live from its equation.
- Cloudflare hosting with a custom domain; Node.js test suites and headless-Chrome tests run
against the live site.
- AI-assisted engineering with Claude Code in VS Code, under written standards.
What it does not claim
- It is a portfolio piece and skills demonstration, not a commercial product: free, with no ads, no cookies
and no sign-in.
- Rules cannot write new sentences: the humanizer rewrites less freely than an AI paraphraser, and the question
tool can miss an answer worded differently from the question.
- The statement reader has been tested on one bank's layout so far. It points out patterns; it is not
financial advice.
13tools
95hostile edge cases
6,000randomized texts, 0 facts changed
0bytes sent while tools run