Private AI engine

The complete DocxIntel AI engine, running on your server.

Most document AI is an API you send pages to. DocxIntel is the opposite: our proprietary OCR, embedding, reranking and reasoning models — and the parsing, intelligence, search, workflow, agent and audit engines around them — are installed inside your data centre or private cloud. Your data stays where you deploy it.

Proprietary models, on your hardware
0 outbound AI requests
Customer-controlled data boundary
Air-gap capable

The DocxIntel AI engine installed inside an isolated customer data centre with no outbound connection

4
Proprietary AI models: 10
Engine components: 0
External AI providers required: 99.0%
Field-level accuracy

Accuracy is measured on the DocxIntel benchmark set with the methodology published on the accuracy page. The figure that binds us commercially is the threshold measured on your own documents in a Proof of Value.

The private AI engine

You don't get an API key. You get the whole engine.

Every model DocxIntel uses — OCR, embedding, reranking and reasoning — is our own, and every one of them is installed on your hardware. Documents travel from your storage to the engine and back, and the path never crosses your perimeter.

Your infrastructure · customer-controlled data boundaryOn-premise, private cloud or air-gapped

ContractsClaim filesKYC packsInvoicesClinical recordsCase files

DocxIntel private AI engine

OCR
Proprietary
Embedding
Proprietary
Retrieval
Private index
Reranking
Proprietary
Reasoning
Proprietary

Document intelligence
Fields, clauses, risks, citations
Workflow & agent
Review, approve, sign, execute
Business action
In your systems, with an audit trail

External AI providers: not required, not called. Outbound AI requests in a default install: 0.

Ten components, one deployment, all on your servers

OCR engineArabic and English print, handwriting, stamps and degraded scans.
Embedding modelSemantic representation of every page, computed locally.
Reranking modelOrders retrieved passages by relevance before reasoning.
Reasoning modelAnswers, drafts, comparisons and risk calls, with citations.
Parsing engineLayout, reading order, tables and multi-document splitting.
Document intelligenceFields, entities and clauses with page and coordinates.
Search & vector storeHybrid keyword and vector retrieval inside your network.
Workflow engineReview, approval, signing and execution steps you define.
Agent engineCarries out multi-step document tasks under your permissions.
Audit & governanceWho did what, to which document, and on what evidence.

The models

Four proprietary models, built for documents in this region

Each model does one job in the pipeline, and each runs on GPUs inside your boundary. None of them falls back to a hosted endpoint when a page is hard or the queue is long.

DocxIntel OCR

Reads the page

Arabic and Latin script, print and handwriting, Arabic-Indic numerals, stamps over text, phone photographs and multi-generation faxes.

Coordinates on every block
Calibrated confidence
RTL and bidirectional reading order

DocxIntel Embed

Understands meaning

Turns every page and passage into a semantic representation, in Arabic and English, so search finds what a clause means and not only the words it uses.

Computed locally
Stored in your vector index
Bilingual retrieval

DocxIntel Rerank

Picks the right evidence

Reorders retrieved passages by real relevance before anything is reasoned over, which is what keeps answers grounded and citations precise.

Fewer, better passages
Higher citation precision
Runs on your GPUs

DocxIntel Reason

Does the work

Answers questions, drafts, compares versions, flags risks and prepares approval packages — every output tied back to the page it came from.

Cited answers
Drafting and comparison
Risk and obligation detection

The engines around them

Models alone do not run document work. These six engines do.

Parsing engine

Layout, reading order, tables, and splitting multi-document packets into the documents they contain.

Document intelligence engine

Fields, entities, clauses and classifications against your own schemas and taxonomy, with page and bounding box on every value.

Search and vector infrastructure

Hybrid keyword and vector retrieval over the collections you choose to index, with access control per collection.

Workflow engine

Review, approval, signing and execution steps you define, with confidence gates that decide what needs a person.

Agent engine

Carries out multi-step document tasks — gather, compare, draft, route — strictly within the permissions of the user it acts for.

Audit and governance layer

Every read, extraction, answer, approval and agent action recorded in a log that lives in your estate, not a vendor portal.

The processing path

From the document arriving to the decision being monitored

Every step below happens inside your environment. The first six are the engine; the rest are your people and your systems, supported by it.

Engine · private ingestion
The document enters your environment
From a watched folder, a mailbox, object storage, your content system or the API — and it stays on storage you control from that moment on.
Engine · proprietary OCR
The page is read
Normalised, deskewed and recognised, with every block carrying its coordinates and a calibrated confidence score.
Engine · proprietary embedding
The content is understood
Passages are embedded locally and written to the vector index inside your environment.
Engine · private retrieval and reranking
The right evidence is selected
Hybrid search finds candidate passages and the reranking model keeps the ones that actually answer the task.
Engine · proprietary reasoning
The work is done
Extraction, comparison, drafting, risk detection or a cited answer — produced on your GPUs.
Engine · AI result
A result with its evidence
Every value and every answer carries its source document, page and region, so it can be checked in seconds.
People · human review
Low-confidence items go to a reviewer
Only what falls below your threshold is queued, with the source region highlighted.
People · approval
An approver decides
On a prepared package of findings and evidence, with the decision written to the audit log.
People · sign
Authorised signatories sign
The signed version is checked against what was approved.
Systems · execute
The result lands in your systems
Structured output and agent actions reach your core systems within the permissions you set.
Systems · monitor
Obligations are tracked
Expiries, renewals and deadlines across the signed estate are watched, and owners are alerted.

The private AI environment

What your administrators see in the AI Infrastructure view

Business users see one indicator: Private AI, running on your infrastructure. Platform and security teams get the full picture.

Deployment

Deployment mode: On-premise, private cloud or air-gapped
Environment: Customer infrastructure
Data residency: Customer controlled

AI engine

Components: OCR, embedding, reasoning, reranking, document intelligence and agent runtime — each with live status
Search: Vector, keyword and hybrid retrieval — each with live status
Model management: Model, purpose, status, version and deployment location for every DocxIntel model

Security

External AI providers: Disabled by default; administrator controlled
Outbound AI requests: 0 in a default deployment
Document data boundary: Customer infrastructure
Audit logging: Enabled, stored in your environment

Detailed infrastructure metrics — GPU utilisation, throughput and queue depth — are exported to the monitoring stack you already run, using the dashboards and alert rules that ship with the deployment.

Why it matters

A private engine versus an AI API

Both can read a document. Only one lets you tell a regulator that the document never left.

A private engine versus an AI API
Criterion	Cloud document AI API	DocxIntel private engine
Where the page is read	Provider infrastructure	Your infrastructure
Where embeddings and indexes live	Usually provider-side	Your vector store
Where prompts and answers are processed	Provider model endpoint	Your GPUs
Outbound AI requests	One or more per page	Zero
AI vendor in your data path	Yes — a processor to assess	No
Works with no internet route	No	Yes, air-gapped by design
Cost of processing more pages	Charged per page or token	No incremental charge

Where the page is read

Cloud document AI API: Provider infrastructure
DocxIntel private engine: Your infrastructure

Where embeddings and indexes live

Cloud document AI API: Usually provider-side
DocxIntel private engine: Your vector store

Where prompts and answers are processed

Cloud document AI API: Provider model endpoint
DocxIntel private engine: Your GPUs

Outbound AI requests

Cloud document AI API: One or more per page
DocxIntel private engine: Zero

AI vendor in your data path

Cloud document AI API: Yes — a processor to assess
DocxIntel private engine: No

Works with no internet route

Cloud document AI API: No
DocxIntel private engine: Yes, air-gapped by design

Cost of processing more pages

Cloud document AI API: Charged per page or token
DocxIntel private engine: No incremental charge

Describes the general architecture of hosted document AI APIs as of 2026. Individual services differ — ask any vendor, including us, for a data-flow diagram and verify it against your own egress logs.

Questions about the private AI engine

What runs where, whose models they are, and what leaves your network.

Are the models really yours, or a wrapper around someone else’s API?

They are ours. The OCR, embedding, reranking and reasoning models are proprietary to BizfyLabs and ship inside the DocxIntel deployment. There is no hosted model behind them: inference happens on the GPUs in your environment, and the engine has no code path that forwards a page, a passage or a prompt to a third-party AI provider.

What exactly leaves our network?

Can an administrator enable an external AI provider anyway?

Are our documents used to train the models?

What hardware does the engine need?

Will business users see infrastructure detail they do not need?

Accuracy benchmarkWhat the engine achieves on hard Arabic and English documents, with the method attached.Deployment modelsAir-gapped, private cloud, managed single-tenant or evaluation sandbox.What's in the boxEvery container, model bundle and artefact installed with the engine.Security postureKey custody, isolation, audit and what we do and do not claim.CapabilitiesAsk, analyse, identify, classify and compare — what the engine does for your teams.Proof of ValueDeploy the engine in your environment for 30 to 45 days and measure it on your documents.

Put the engine on your own server and test it.

A fixed-fee Proof of Value deploys the complete DocxIntel engine in your environment, measures accuracy on your documents and lets your security team watch the egress logs while it runs.

Runs in your environment
Zero outbound AI requests
You keep the report either way

Private AI engine

The complete DocxIntel AI engine, running on your server.

Proprietary models, on your hardware
0 outbound AI requests
Customer-controlled data boundary
Air-gap capable

4
Proprietary AI models: 10
Engine components: 0
External AI providers required: 99.0%
Field-level accuracy

The private AI engine

You don't get an API key. You get the whole engine.

Your infrastructure · customer-controlled data boundaryOn-premise, private cloud or air-gapped

ContractsClaim filesKYC packsInvoicesClinical recordsCase files

DocxIntel private AI engine

OCR
Proprietary
Embedding
Proprietary
Retrieval
Private index
Reranking
Proprietary
Reasoning
Proprietary

Document intelligence
Fields, clauses, risks, citations
Workflow & agent
Review, approve, sign, execute
Business action
In your systems, with an audit trail

External AI providers: not required, not called. Outbound AI requests in a default install: 0.

Ten components, one deployment, all on your servers

OCR engineArabic and English print, handwriting, stamps and degraded scans.
Embedding modelSemantic representation of every page, computed locally.
Reranking modelOrders retrieved passages by relevance before reasoning.
Reasoning modelAnswers, drafts, comparisons and risk calls, with citations.
Parsing engineLayout, reading order, tables and multi-document splitting.
Document intelligenceFields, entities and clauses with page and coordinates.
Search & vector storeHybrid keyword and vector retrieval inside your network.
Workflow engineReview, approval, signing and execution steps you define.
Agent engineCarries out multi-step document tasks under your permissions.
Audit & governanceWho did what, to which document, and on what evidence.

The models

Four proprietary models, built for documents in this region

Each model does one job in the pipeline, and each runs on GPUs inside your boundary. None of them falls back to a hosted endpoint when a page is hard or the queue is long.

DocxIntel OCR

Reads the page

Arabic and Latin script, print and handwriting, Arabic-Indic numerals, stamps over text, phone photographs and multi-generation faxes.

Coordinates on every block
Calibrated confidence
RTL and bidirectional reading order

DocxIntel Embed

Understands meaning

Turns every page and passage into a semantic representation, in Arabic and English, so search finds what a clause means and not only the words it uses.

Computed locally
Stored in your vector index
Bilingual retrieval

DocxIntel Rerank

Picks the right evidence

Reorders retrieved passages by real relevance before anything is reasoned over, which is what keeps answers grounded and citations precise.

Fewer, better passages
Higher citation precision
Runs on your GPUs

DocxIntel Reason

Does the work

Answers questions, drafts, compares versions, flags risks and prepares approval packages — every output tied back to the page it came from.

Cited answers
Drafting and comparison
Risk and obligation detection

The engines around them

Models alone do not run document work. These six engines do.

Parsing engine

Layout, reading order, tables, and splitting multi-document packets into the documents they contain.

Document intelligence engine

Fields, entities, clauses and classifications against your own schemas and taxonomy, with page and bounding box on every value.

Search and vector infrastructure

Hybrid keyword and vector retrieval over the collections you choose to index, with access control per collection.

Workflow engine

Review, approval, signing and execution steps you define, with confidence gates that decide what needs a person.

Agent engine

Carries out multi-step document tasks — gather, compare, draft, route — strictly within the permissions of the user it acts for.

Audit and governance layer

Every read, extraction, answer, approval and agent action recorded in a log that lives in your estate, not a vendor portal.

The processing path

From the document arriving to the decision being monitored

Every step below happens inside your environment. The first six are the engine; the rest are your people and your systems, supported by it.

Engine · private ingestion
The document enters your environment
From a watched folder, a mailbox, object storage, your content system or the API — and it stays on storage you control from that moment on.
Engine · proprietary OCR
The page is read
Normalised, deskewed and recognised, with every block carrying its coordinates and a calibrated confidence score.
Engine · proprietary embedding
The content is understood
Passages are embedded locally and written to the vector index inside your environment.
Engine · private retrieval and reranking
The right evidence is selected
Hybrid search finds candidate passages and the reranking model keeps the ones that actually answer the task.
Engine · proprietary reasoning
The work is done
Extraction, comparison, drafting, risk detection or a cited answer — produced on your GPUs.
Engine · AI result
A result with its evidence
Every value and every answer carries its source document, page and region, so it can be checked in seconds.
People · human review
Low-confidence items go to a reviewer
Only what falls below your threshold is queued, with the source region highlighted.
People · approval
An approver decides
On a prepared package of findings and evidence, with the decision written to the audit log.
People · sign
Authorised signatories sign
The signed version is checked against what was approved.
Systems · execute
The result lands in your systems
Structured output and agent actions reach your core systems within the permissions you set.
Systems · monitor
Obligations are tracked
Expiries, renewals and deadlines across the signed estate are watched, and owners are alerted.

The private AI environment

What your administrators see in the AI Infrastructure view

Business users see one indicator: Private AI, running on your infrastructure. Platform and security teams get the full picture.

Deployment

Deployment mode: On-premise, private cloud or air-gapped
Environment: Customer infrastructure
Data residency: Customer controlled

AI engine

Components: OCR, embedding, reasoning, reranking, document intelligence and agent runtime — each with live status
Search: Vector, keyword and hybrid retrieval — each with live status
Model management: Model, purpose, status, version and deployment location for every DocxIntel model

Security

External AI providers: Disabled by default; administrator controlled
Outbound AI requests: 0 in a default deployment
Document data boundary: Customer infrastructure
Audit logging: Enabled, stored in your environment

Why it matters

A private engine versus an AI API

Both can read a document. Only one lets you tell a regulator that the document never left.

A private engine versus an AI API
Criterion	Cloud document AI API	DocxIntel private engine
Where the page is read	Provider infrastructure	Your infrastructure
Where embeddings and indexes live	Usually provider-side	Your vector store
Where prompts and answers are processed	Provider model endpoint	Your GPUs
Outbound AI requests	One or more per page	Zero
AI vendor in your data path	Yes — a processor to assess	No
Works with no internet route	No	Yes, air-gapped by design
Cost of processing more pages	Charged per page or token	No incremental charge

Where the page is read

Cloud document AI API: Provider infrastructure
DocxIntel private engine: Your infrastructure

Where embeddings and indexes live

Cloud document AI API: Usually provider-side
DocxIntel private engine: Your vector store

Where prompts and answers are processed

Cloud document AI API: Provider model endpoint
DocxIntel private engine: Your GPUs

Outbound AI requests

Cloud document AI API: One or more per page
DocxIntel private engine: Zero

AI vendor in your data path

Cloud document AI API: Yes — a processor to assess
DocxIntel private engine: No

Works with no internet route

Cloud document AI API: No
DocxIntel private engine: Yes, air-gapped by design

Cost of processing more pages

Cloud document AI API: Charged per page or token
DocxIntel private engine: No incremental charge

Questions about the private AI engine

What runs where, whose models they are, and what leaves your network.

Are the models really yours, or a wrapper around someone else’s API?

What exactly leaves our network?

Can an administrator enable an external AI provider anyway?

Are our documents used to train the models?

What hardware does the engine need?

Will business users see infrastructure detail they do not need?

Put the engine on your own server and test it.

A fixed-fee Proof of Value deploys the complete DocxIntel engine in your environment, measures accuracy on your documents and lets your security team watch the egress logs while it runs.

Runs in your environment
Zero outbound AI requests
You keep the report either way

The complete DocxIntel AI engine, running on your server.

You don't get an API key. You get the whole engine.

Ten components, one deployment, all on your servers

Four proprietary models, built for documents in this region

Reads the page

Understands meaning

Picks the right evidence

Does the work

Models alone do not run document work. These six engines do.

Parsing engine

Document intelligence engine

Search and vector infrastructure

Workflow engine

Agent engine

Audit and governance layer

From the document arriving to the decision being monitored

The document enters your environment

The page is read

The content is understood

The right evidence is selected

The work is done

A result with its evidence

Low-confidence items go to a reviewer

An approver decides

Authorised signatories sign

The result lands in your systems

Obligations are tracked

What your administrators see in the AI Infrastructure view

Deployment

AI engine

Security

A private engine versus an AI API

Questions about the private AI engine

Keep reading

Put the engine on your own server and test it.

The complete DocxIntel AI engine, running on your server.

You don't get an API key. You get the whole engine.

Ten components, one deployment, all on your servers

Four proprietary models, built for documents in this region

Reads the page

Understands meaning

Picks the right evidence

Does the work

Models alone do not run document work. These six engines do.

Parsing engine

Document intelligence engine

Search and vector infrastructure

Workflow engine

Agent engine

Audit and governance layer

From the document arriving to the decision being monitored

The document enters your environment

The page is read

The content is understood

The right evidence is selected

The work is done

A result with its evidence

Low-confidence items go to a reviewer

An approver decides

Authorised signatories sign

The result lands in your systems

Obligations are tracked

What your administrators see in the AI Infrastructure view

Deployment

AI engine

Security

A private engine versus an AI API

Questions about the private AI engine

Keep reading

Put the engine on your own server and test it.