← All features

Every row points back at the line it came from

A bookkeeper signing off 180 rows needs to check the ones that look wrong against what the bank actually printed. Doing that in a separate PDF viewer, scrolling to find the line, is enough friction that most people skip it and hope.

How it works on a digital statement

A PDF with a text layer records where every piece of text sits on the page. The parser keeps that position as it builds each transaction, so a row carries the coordinates of the line it was read from. Selecting the row scrolls to the page and draws a band over the line.

Where a row came through the model path rather than the geometric parser, the line is found by searching the page’s own text for the row’s words — the same way a person would, but instantly.

How it works on a scan, which is the hard case

A scanned page has no text to measure. The lines are found in the pixels instead: the page is rendered to a canvas, and rows of the image carrying text are dark where the gaps between them are not.

Two things printed on a ruled statement are not text and had to be handled before any of it worked. A horizontal rule reads as a very dark line and welds the rows above and below it into one band. A vertical rule — a table’s right-hand border — sits further right than every amount, so on every line it becomes "the rightmost mark", which quietly destroys the one content signal a scan offers. Both are found the same way: a printed rule runs unbroken, and text does not.

The estimate is then corrected against content. Amounts are right-aligned, so a statement’s transaction lines share a right edge that headings, addresses and totals do not — which is what keeps a row from being highlighted on the payment slip at the foot of the page, however far out the estimate was.

Why this is worth the engineering

It is the difference between a tool you check and a tool you trust. Reconciliation tells you row 47 is wrong; this shows you what row 47 says on the page. Without it, every flagged row costs a trip to a PDF viewer, and after three of those most people stop looking.

On a poor scan the band is an estimate

Where a page is faint, skewed, or annotated by hand, the line can only be located approximately, and the interface says so rather than drawing a confident band in the wrong place. You are told whether the highlight is exact, located from the page image, or an estimate — because a band that looks certain and is wrong is worse than one that admits it is unsure.

Related