Kiliyo

Unstructured PDF to structured content

Your PDF, structured. Every word kept.

Kiliyo reads the geometry of the printed page, not just its text, and rebuilds it as clean, structured content: headings, sections, notes, quotations, tables, index. Hours instead of days, and a quality scorecard with every document.

Get a free sample on your own PDF See how it works

Born-digital PDFs, any length. Already in production for documents in English, Danish, Icelandic and Welsh.

One structure, many outputs

Structure it once. Use it wherever you need it.

Kiliyo first rebuilds what the page means: headings, paragraphs, notes, tables and reading order. Every output is built from that one structure.

Available now

Reflowable reading

Text that adapts to any screen, with live contents, notes and index. Delivered today in four languages.

In development

Markdown

Clean text with real headings, for docs sites and knowledge bases.

In development

Structured JSON

Every block with its role and reading order, ready for search and AI pipelines.

In development

Tables to spreadsheets

Ruled tables lifted out as CSV, row for row.

Measured, not promised

The numbers we put in front of every client

Every conversion ships with a QA scorecard that compares the output against the source PDF word by word and paragraph by paragraph. These are the figures it reports on real documents in production today.

98–100%
of words carried over exactly. The engine never rewrites text.
~80%
of structure right on the first pass (sections, notes, quotations, index). The scorecard shows where review is needed.
<2min
of machine time for a 300-page document. Your turnaround becomes a review step, not a rebuild.
0
invented words. Deterministic geometry, with optional AI assistance that only reformats, never rewrites.

Checked, not trusted

Your words go in. The same words come out.

A file can pass every validator and still be wrong. So Kiliyo never takes its own work on trust: the engine and optional AI propose, a strict check decides, and anything uncertain goes to a person.

1ProposePage geometry, with optional AI, suggests the structure: joins, headings, quotations.
2CheckEvery AI fix is compared with your source text. One changed word and it’s rejected.
3ReviewWhat the check can’t settle is listed for a person, paragraph by paragraph.

Don’t trust the clever part. Check its work.

The same rule keeps flight-control software, theorem provers and DNA sequencers honest.
In developmentA confidence score on every paragraph, showing how sure the check is about each one that passed, not only which ones failed.

How it works

Five steps from PDF to structured content

Sections are detected automatically. A person confirms the front and back matter, which takes minutes, and the engine does the rest.

Upload

Drop the PDF. Title, author, identifiers and language are captured as metadata.

Detect

Sections are found from the printed contents page and heading geometry, with a confidence score per page.

Confirm

Cover, title, contents, notes, appendices, index: tagged in a visual page map in a few minutes.

Build

Paragraphs are reflowed across page breaks; footnotes, endnotes, block quotations and the index become live links.

Check

The QA scorecard compares output to source and lists the exact paragraphs to review. Fix by hand or with AI assist, in the browser.

Fit

What it handles, and what it doesn't

We would rather you know before you send a document than after.

Built for

  • Born-digital print PDFs from InDesign, LaTeX or similar, any page count.
  • Long-form text: monographs, reports, non-fiction, fiction and poetry.
  • Front and back matter reproduced faithfully: contents pages, epigraphs, dedications, prefaces, acknowledgements.
  • Footnotes, endnotes and indexes turned into working links, including two-column indexes.
  • Verse and block quotations kept as set-off blocks with their indents.
  • Images and simple ruled tables extracted and placed.
  • Non-English text in Latin scripts. Danish, Icelandic and Welsh documents already delivered.

Not a fit today

  • Scanned or image-only PDFs. There is no OCR step; the text layer must exist.
  • Fixed-layout designs: picture-led pages, comics, heavily designed spreads.
  • Maths-heavy or code-heavy texts. Equations and code listings are not preserved as such.
  • Right-to-left scripts.
  • Complex borderless tables. Simple ruled tables work; dense data tables need manual work.

Free sample, no strings

Send us one PDF. Get it back structured, with its scorecard, within a day.

Email the PDF, or a link to it, and we return the structured output plus the QA report so you can judge it on your own document rather than ours. If it isn't good enough, you've lost nothing.

hk24@kiliyo.com Email a PDF
  • Any length, any Latin-script language.
  • Files are processed on our own server, deleted within two hours of the conversion, and never used for anything else.
  • You get the structured output, the QA scorecard and a plain note on what would need review for a full run.
  • Typical reply: within one working day.