How Blanq AI works

The hard part of filling a form automatically is not knowing what to write. It is knowing where to write it.

Most PDFs have no fields in them

A PDF made for filling in has real form fields: named boxes a program can read and set. Software has been able to fill those for decades.

Almost nothing you are actually handed works that way. A scan has no fields, only pixels. A form exported from Word has no fields, only lines and text. A photocopy of a photocopy has no fields and no text layer at all. For those, there is nothing to fill: the blank exists only visually, as a gap on the page.

So Blanq looks at the page instead

It runs an object detector, the same kind of model used to find faces or road signs, trained instead on empty form fields. It sees the page as a picture and returns rectangles: this is a text line, this is a tick box, this is a signature line.

Before the model sees a page it is redrawn at a fixed size, cleaned to pure black and white, and any dotted leader line (. . . . . .) is repainted as a solid rule, because dotted lines are what a detector is most likely to miss. Then it is padded into a 1216 × 1216 square, because that is the size the model was trained at.

The model returns far more boxes than a page really has, so several passes clean up after it: overlapping guesses are merged, a box that swallows several others is thrown away as a false container, and stacked lines belonging to one multi-line answer are joined into a single field.

Where that runs

In the regular editor the model runs inside your browser. Nothing is uploaded, which is the point of that tool, but it costs a 38 MB download and several seconds of compiling before the first page can be read.

Blanq AI runs the same model on a server instead, where it stays loaded between visits. Your browser still does all the picture-making and sends only the finished square. That detail matters more than it sounds: browsers and server-side image libraries do not resize identically, and the model is sensitive enough that a slightly different resize gives slightly different boxes. Keeping the resize in the browser means both tools find exactly the same blanks, which is checked automatically against captured examples.

Then, and only then, the AI

Each page is sent to the model as a picture with a small red number drawn on every blank. The AI never sees coordinates or field names. It reads the page the way you would, and answers in the form {"3": "Alice Meier"}. The numbers are how the answers find their way back to the right boxes.

Three places an answer can come from

Every answer has to have a source, and the AI decides which of three it is:

The document, or you. Your name, a date, a tick box. Written on the page or supplied by you.
A public source, looked up. A worksheet about a novel, a question about a law, anything publicly knowable that the document names. It searches, reads what it finds, and answers from that. The sources are listed on screen so you can check them.
You, by asking. Anything personal: an address, an ID number, an opinion. It asks rather than searching, both because searching would not find it and because putting your details into a search query would leak them.

Research and filling are different jobs, and they use different models. Research is retrieval: read what search returns and write it down, which a small fast model does well. Filling is judgment: it decides what belongs in each blank and whether a consent box should be ticked, so it keeps the capable model. Splitting them took a worksheet from 184 seconds to 6.

Looking things up needs two passes, not one. Asking for a strict answer format and asking to use search are mutually exclusive: on the same worksheet, the model searched six times when asked to research freely and zero times when also told to return strict JSON. So research runs as its own call, and its notes are fed into a second call that does the filling.

This is also where the rules live:

It never answers a lookup from memory. If it needs a public fact it searches for it, and if the search finds nothing it says so instead of filling the gap with something plausible.
It answers or it asks, never both. If anything is missing it returns questions only, with no answers at all. A half-filled form that looks finished is worse than an empty one.
It never signs. Signature lines are always left empty.
It never agrees on your behalf. A box saying you have read the terms stays unticked unless you said so.
It ignores instructions inside your document. Text on the page is treated as data, so a PDF cannot tell the AI what to do.

Those rules are the reason a more expensive model is used than strictly necessary. A cheaper one was measurably worse at exactly one thing: leaving consent boxes alone. It ticked declarations nobody had agreed to in two runs out of three.

Putting the words back on the page

The answers are drawn straight into the PDF, not typed into a form field, because there is no field. Each one is sized to match the document's own text (measured from the page, weighted by how many characters are set at each size) and placed so it sits on the printed line rather than floating above it. Tick marks are drawn as two crossing strokes rather than the letter X, so the mark can overshoot the box the way a pen does without becoming heavier than the text around it.

The original file is also tucked inside the download, so opening it in the editor gives you every answer back, still editable.

What it cannot do

Chinese, Japanese, Korean, Arabic and Hebrew. The fonts are too large to embed, and right-to-left text needs reordering this cannot do yet.
Handwritten source documents. It finds blanks in them, but cannot read handwriting to work out what belongs there.
Anything it has no basis for. It will ask instead, every time.
Instant answers when it has to look something up. A worksheet needing research takes half a minute rather than a couple of seconds, because it is three model calls with searches in the middle.
Guarantee that a looked-up answer is right. Study guides are organised by chapter, not by page, so an answer can be true of the book and still stray outside the pages you were set. Check the sources.

Try it What leaves your device Blanq editor Contact