DocxIntel installs inside infrastructure you already control — a disconnected data centre, your own VPC, a dedicated single-tenant stack, or a time-boxed sandbox. The container images, the model weights and the API surface are identical across all four. Only the operating boundary changes.
Install windows assume the platform prerequisites below are already met — Kubernetes or Docker host provisioned, GPU drivers loaded, storage classes and ingress in place. Prerequisite work is scoped in writing before an install date is agreed.
Every model runs the same software on the same open-weight models. What differs is who operates the stack, where the perimeter sits, and how updates cross it.
A fully disconnected install inside your own data centre. No route to the internet, no proxy, no DNS beyond your own zones. Updates arrive as a signed offline bundle that you verify and import on your terms.
DocxIntel deployed into your own AWS, Azure or OpenStack tenancy — your VPC, your subnets, your KMS. Private endpoints keep object storage and identity traffic off the public internet, and internet egress can be left switched off entirely.
A dedicated stack — dedicated GPUs, dedicated database, dedicated model serving — operated day to day by BizfyLabs engineers under access you grant and can revoke. Nothing is shared with another customer at any layer.
A single-node, time-boxed install for teams who want to put their own documents through the platform before committing to anything. One GPU, disposable storage, and a defined end date after which the environment is destroyed.
Most document AI platforms treat disconnection as an exception to be engineered around. DocxIntel is built the other way: the offline path is the primary path, and the connected deployments are the ones with extra convenience layered on top.
An air-gapped install performs identical analysis, extraction and question answering to a connected one. There is no degraded offline mode because there is no online mode to degrade from.
When the constraint is data residency and tenancy rather than physical disconnection, DocxIntel deploys into your own cloud account and stays inside it. The cluster runs in private subnets, reaches object storage over gateway or private endpoints, and needs no NAT gateway to function.
Deploying inside your own cloud tenancy is data residency. It is not the same as air-gapped operation, and we will not describe it that way — if your policy requires physical disconnection, pick Model 01.
Some organisations want the platform without the platform team. Managed single-tenant gives you a dedicated stack that BizfyLabs engineers operate — patching, upgrades, capacity, on-call — while custody of the data and the keys stays with you.
Managed does not mean multi-tenant. Every operator action runs through your identity provider, is scoped to a named engineer, and is recorded in an audit trail you can export.
The fastest way to find out whether DocxIntel reads your documents is to run your documents through it. The sandbox is a single-node install designed to be stood up quickly, used hard for a few weeks and then destroyed.
Sandbox results carry forward. The extraction schemas and classification taxonomies you build during evaluation import directly into the Proof of Value and then into production.
The software is the same in every column. What changes is the perimeter, the operator and the path an update takes to reach you.
Network egress
Model weights location
Update mechanism
Key custody
Typical time to deploy
Best for
Licensing does not change between models. DocxIntel is a fixed annual licence sized by deployment footprint — GPU count and environment count — never by pages processed, so re-running your archive after a model update costs nothing.
From first architecture call to production sign-off. The sequence is the same in every model; only the duration of step 4 changes.
One session with your infrastructure, security and data teams. We agree the deployment model, the target cluster, GPU class and count, storage classes, ingress, identity integration and the document intake path. The output is a written prerequisite checklist, not a proposal deck.
Nodes provisioned, NVIDIA drivers and container toolkit installed, storage classes created, ingress and certificates in place, OIDC or SAML client registered, and the internal registry ready to receive images. Nothing here is DocxIntel-specific — it is standard platform work your team already knows.
You receive the release bundle — OCI image archives, model weights, Helm chart, SBOM and a SHA-256 manifest with a cosign signature. You verify it, scan it with your own tooling, and load it into your registry. In air-gapped sites this arrives on encrypted media or through your one-way transfer gateway.
A single Helm release brings up the gateway, orchestrator, GPU services, worker pool and admin console. We connect your identity provider, wire the intake path, run the built-in health and GPU validation suite, and process a known document set end to end.
Extraction schemas, classification taxonomies and confidence thresholds are tuned on your real document mix, including the difficult tail — phone photos, handwritten Arabic, stamped contracts and multi-generation faxes. Field-level accuracy is measured against a human-adjudicated ground truth.
Runbooks, dashboards, alert thresholds, backup and restore drills, and an upgrade rehearsal. Your team performs the first upgrade themselves with us alongside, so the second one does not need us at all.
What needs to exist before install day. These are the same in all four deployment models, and every one of them is standard platform infrastructure rather than something DocxIntel invents.
Exact figures for your environment are confirmed against your document mix and volume during the architecture review, and again on your own hardware during the Proof of Value. Installation reference material is maintained in the product documentation.
The questions that decide whether a deployment gets approved, answered without hedging.
Neither. Air-gapped is the default assumption the product is built around, and it is available on every licence. There is no premium SKU, no enterprise gate and no separate air-gap edition — the same signed bundle installs in a disconnected data centre and in a connected VPC. Vendors who gate self-hosting behind their top tier are describing a different product shape.
