Turn any document into
structured, verified data.

Flux reads the document with whichever engine handles it best, then checks its own output before it hands it to you.

Request access

Book a demo

READS

PDF

Scans

Photographs

Spreadsheets

Word files

Handwriting

Account holder

Trade and Industry Co.

0.99

Tax number

4820561973

1.00

Requested limit

500000.00

0.97

Authorised signatory

[redacted]

0.51

1 field sent for review

Read anything, with any engine.

No single model reads every document well. Flux picks per document, audits what comes back, and tells you how much of it to trust.

SCANNED · SKEWED · MIXED CONTENT

Engine A 0.71ROUTED to engine_b · 0.91Engine C 0.64

PARSED OUTPUT

## Signature Authorisation Form

No. 4471902

| [redacted] | Managing Dir.    |

| [redacted] | Finance Director |

| [redacted] | Operations Mgr.  |

### Signatures

3 signatures detected

Reading engines, vision models and layout models each handle different documents better. Flux routes each document to the engine most likely to read it correctly, and new engines register without a code change.

On the settings that call for it, several engines read the same page at once. They do not see each other's output, so agreement between them means something.

The readings are scored against each other, text that repeats is penalised, and on the top setting a model that can see the page checks the winner against the original image.

The surviving text becomes the fields you defined, with a confidence signal on each one and an empty value where nothing could be read.

Keep scrolling

Fast

One engine reads the page, quickest result, no cross check by design.

Reliablerecommended

Several engines read the same page and the strongest reading wins with the reason recorded.

Premiumexperimental

A model that can see the page checks the winner against the original image.

WHAT IT HANDLES

The documents that break other pipelines.

Handwriting

Filled forms, margin notes, endorsements on the back of a cheque, read and returned as fields, not as an image.

PARSED

Ref: LN-2026-004182

Borrower: [redacted]

Amount: 1,250,000

Term: 36 months

Rate: 42.00%

handwriting resolved

Tables

Dense, irregular, spanning pages, with merged cells and no ruling. Rows and relationships come back intact.

STRUCTURED

name         authority

[redacted]   Sole

[redacted]   Joint

[redacted]   Joint

merged cells resolved

Signatures

Located, cropped and verified against an enrolled signer, with the score and the coordinates behind the verdict.

sig 1 · 0.96

VERDICT

vs. [redacted]0.94

Match

cluster cohesion 0.97
page 2 · [46,132,202,180]

Scans that arrived badly

Rotation, skew and low contrast corrected before reading, and flagged when the scan is too poor to trust.

low contrast

deskewed · readable

OUR VIEW

The market broke twice.

Document processing has been rebuilt twice. Neither rebuild solved the thing that actually blocks automation.

Templates broke first

OCR could only read what someone had already mapped. Every new document type was a project, and anything that arrived off template failed silently, or returned the wrong field with no sign anything was wrong.

Then models broke it differently

Language models read anything you give them. They also invent, fluently and without warning. Flexibility arrived, and with it a system that sounds equally certain whether it read the number or made it up.

What both left behind

Teams that can process any document but cannot prove any answer, so a person checks everything anyway, and the automation was never really automation.

CONFIDENCE

Know when it’s right, and when it isn’t.

Every result carries the evidence behind it, so your workflow knows what to trust, and what to send to a person.

Extraction resultReview required

FIELDVALUECONFIDENCE

Account holderTrade and Industry Co.

0.99

Tax number4820561973

1.00

Requested limit500,000.00

0.97

Authorised signatory[redacted]

0.51

Source: page 1, [412, 1180, 664, 1262]

Field-level confidence

One score per extracted value, not one for the page.

Quality assessment

Completeness, accuracy and coherence, plus a penalty when output drifts from the source.

Source grounding

Page and coordinates for every value, so anything can be traced back.

A number you cannot audit is not a result. It is a guess with formatting.

CAPABILITIES

One pipeline. Six things it can be asked to do.

Each one is a call, and each one leaves a record of how it decided.

Orbina Flux

Flux reads a page with whichever engine handles it best, checks its own output, and returns fields with a confidence signal on each one. It runs on your infrastructure, with your models.

Parsing results

Reads scans, photographs, spreadsheets and word files into clean structured text, keeping the layout, the tables and the handwriting rather than flattening them. Every page is upscaled, sharpened, straightened and stripped of borders before anything reads it.

Runs the page through more than one engine, scores the readings against each other, penalises text that repeats, and leaves a field empty rather than guessing at it.

Turns the text into the fields you defined. Describe them in a sentence and the schema is written for you, versioned in a registry rather than pasted into a prompt.

Cuts a bundle into the documents inside it. A single scan of forty pages becomes the eleven files it actually holds, each with its own type and its own trace.

Sorts a document into a type when nobody named one, using rules written in plain language rather than a training set you have to collect first.

Chunks and embeds the archive so it can be searched by meaning rather than by keyword, and opened to your own systems through a token you control.

THE API

Six calls. Every document.

Every call has a synchronous, asynchronous and batch form. Webhooks on completion. Schema and prompt registries, so extraction logic is versioned rather than pasted.

/parseLayout aware read of any supported file into clean markdown

/extractSchema based extraction, you define the fields, no training

/extract/agenticMulti step extraction for documents that need reasoning across pages

/discoverFind what is in a document set before you know what to ask

/vision/readDirect visual read for images and screenshots

/signatures/verifyLocate signatures and verify them against enrolled signers

Base URL is your own deployment. Endpoints and payloads are identical.

cURL

curl -X POST "https://your-deployment/v1/extract" \
  -H "Authorization: Bearer $FLUX_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
        "file_url": "https://your-storage/statement.pdf",
        "schema_id": "sch_8f21c04b",
        "mode": "reliable"
      }'

Python

from flux import Flux
 
client = Flux(base_url="https://your-deployment")
 
result = client.extract(
    file_url="https://your-storage/statement.pdf",
    schema_id="sch_8f21c04b",
    mode="reliable",
)
 
for field in result.fields:
    print(field.name, field.value, field.confidence)

Fair under load

Throttling steps down as the queue fills, so a large batch cannot starve a small one.

Enterprise grade security

Granular access control, encryption in transit and at rest, and the certifications a regulated buyer asks for.

Dedicated support

Named engineers, response times in writing, and a path to a person who knows your deployment.

Flexible deployment

Run it in our cloud or entirely inside your own, wherever data residence requires.

Enterprise ready

Built for teams running production grade document work, with reliability, security and control at scale.

Book a demo

TRUST

Regulated by default, fast by design.

Built for banks and insurers, without slowing down the teams that ship.

Four deployment modes

Global cloud, local cloud, on-premises or LLM as a Service. Same architecture, your choice.

Data stays where you say

Documents, conversations and logs live in your perimeter. Nothing leaves without a rule that says it can.

ISO 27001 · 42001

Information security and AI management: audited, not self-declared.

Your models, your keys

Bring your own model and keys, or run ours behind your firewall. Swap either without touching the pipeline.

Deployment

Application

Orchestration

Model runtime

Keys and data

Managed

APPLICATIONS

Applications running on Flux today.

Runs on Flux

Cheque Processing

Every cheque read, verified and locked before it reaches the system that pays it.

See application How it works

Extracted fields

AmountHigh

Due dateHigh

Cheque noHigh

DrawerReview

Runs on Flux

Signature Verification

Every signature detected, compared and decided, with a person where the evidence is thin.

See application How it works

Reference

On document

InconclusiveSend to review

Also on Flux: Invoices · Insurance policies · Receipts · Contracts

Start with one concrete workflow.

A document flow, a conversation line, an approval process. Live in weeks.

Book a demo

Meet CXOS →

Meet Flux →

Turn any document intostructured, verified data.

Turn any document intostructured, verified data.

Turn any document intostructured, verified data.

Read anything, with any engine.

The documents that break other pipelines.

The documents that break other pipelines.

The documents that break other pipelines.

Handwriting

Tables

Signatures

Scans that arrived badly

The market broke twice.

The market broke twice.

The market broke twice.

Know when it’s right, and when it isn’t.

Know when it’s right, and when it isn’t.

Know when it’s right, and when it isn’t.

One pipeline. Six things it can be asked to do.

One pipeline. Six things it can be asked to do.

One pipeline. Six things it can be asked to do.

Six calls. Every document.

Six calls. Every document.

Six calls. Every document.

Enterprise ready

Regulated by default, fast by design.

Regulated by default, fast by design.

Regulated by default, fast by design.

Four deployment modes

Data stays where you say

ISO 27001 · 42001

Your models, your keys

Applications running on Flux today.

Applications running on Flux today.

Applications running on Flux today.

Start with one concrete workflow.

Turn any document into
structured, verified data.

Turn any document into
structured, verified data.

Turn any document into
structured, verified data.