Verified extraction

Nothing lost
in translation

We turn PDFs, scans and office files into structured Markdown and JSON — and check every field against the source. On large, capable models that run in-house. Your documents never leave your control.

Talk to us ↓

The problem AI reads documents.
Who checks what it missed?

Converting a document with AI is easy. Knowing that the result is complete is not. The failures are silent — the output looks finished either way.

  • A table loses its last rows.
  • A long document is cut off mid-way.
  • Footnotes, annexes and small print disappear.
  • Two columns merge, a number shifts by one line.
  • A value is “corrected” that was never in the source.
SOURCE · p. 14 Pos 1 12'400.00 Pos 2 3'150.00 Pos 3 880.50 Pos 4 7'200.00 Total 23'630.50 EXTRACTED {"pos":1,"amt":12400} {"pos":2,"amt":3150} {"pos":3,"amt":880.5} missing · p.14 r.4 "total": 23630.5 Σ rows ≠ total ✕

How it works Every field, checked
against the source

Not samples. Everything.

DOCUMENT PDF · scan · office EXTRACT AI + independent baseline RECONCILE provenance · totals · 2nd pass REPORT MD · JSON · audit trail
01

Independent baseline

We read the document a second way, without AI: text layer, OCR with confidence scores, page and table counts. This is the yardstick.

02

Provenance for every element

Each value in the JSON points to its page and position in the original. Anything in the source without a match is flagged.

03

Control totals

Every number in the document must appear in the output. Row counts, sums and headings must reconcile.

04

Literal grounding

Extracted values must be found verbatim in the source. Invented or “improved” values fail.

05

Two independent passes

A second model extracts again. Differences go to review, not into your data.

06

Verification report

Coverage per page, flagged items, full audit trail. You don't have to trust the result — you can check it.

The dilemma Thorough checks need big models.
Big models usually live in someone else's cloud.

Checking everything takes the most capable models available. Today those run at a handful of providers abroad. Using them means sending exactly the documents you want to protect — contracts, records, designs — out of the house.

Frontier models in the cloud
  • Deep checks — but your documents leave the house.
  • Data under foreign jurisdiction.
  • Or: keep control and check less.
HULLA in-house
  • Large, capable open-weight models on dedicated infrastructure.
  • Same depth of verification, no data leaving your control.
  • If required, fully offline.

Your jurisdiction, your choice Frontier-grade checks. Your data stays home.

Hosted at HULLA in Liechtenstein, Switzerland or Germany — or inside your own environment.

Liechtenstein Switzerland Germany Your premises Air-gapped

Engagement Advise. Build. Operate.

Start with advice. Grow into build or operate when it makes sense.

Advise

Design the setup

We assess your documents and requirements: model choice, hardware sizing, verification pipeline, governance and EU AI Act readiness. For teams who want to run it themselves.

Build

Install on your side

We install and hand over the full pipeline on your infrastructure — on-premise, in your private cloud or fully air-gapped. For data that must stay physically with you.

Operate

Run it with us

We process and verify your documents on HULLA infrastructure in Liechtenstein, Switzerland or Germany — you choose where your data lives. No GPUs to buy, no models to maintain.

Industries Wherever documents are confidential —
and nothing may be missing

And anywhere else your documents should not leave the house.

Banking & wealth managementKYC files, statements, contracts
InsurancePolicies, claims, medical reports
Fiduciary & legalContracts, company records, case files
Industry & engineeringSpecifications, test reports, supplier contracts
Pharma & medtechStudies, regulatory dossiers, batch records
HealthcareFindings, discharge reports, lab results
Security & defenceTechnical documentation, fully offline operation
Public administrationApplications, registers, archives
Energy & infrastructureMaintenance logs, network plans, safety records
Audit & taxFinancial statements, receipts, filings

Questions What people ask us

Find out what your
documents are hiding

Send us a sample. We'll show you what a standard AI conversion misses — and what a verified one looks like.

Send a sample →