PRD · BUDBOXOnDraft

Local AI appliance

A dedicated machine for private, local AI.

On your premises. Under your control.

Schools, businesses and organisations handling pupil records, client files or regulated data can use AI without sending the raw document out of the building. Budbox desensitises on the box, can run fully local models, and only reaches allowlisted AI endpoints when you choose.

Sleek compact Budbox AI appliance on a desk under soft studio lighting
HOW · 01–05On premises

How Budbox works

A dedicated privacy appliance in your building — not another browser tab that uploads the file as-is.

  1. 1

    Install on your premises

    A locked-down AI stack runs in Docker on a dedicated Budbox in your building — not a shared public chat product.

  2. 2

    Airlock sensitive documents

    Before any optional cloud AI call, documents are desensitised on the box. Responses are rehydrated locally. Fully local models stay on the box entirely.

  3. 3

    Enrich with your domain graph

    A local graph database holds organisation knowledge so answers stay grounded in your context, not generic web text.

  4. 4

    Wire into Dexl workflows

    Budbox appears as a special node type in Dexl, including conditional question steps for admin and process flows.

  5. 5

    Egress stays locked down

    Network egress only reaches allowlisted AI endpoints. Everything else stays inside your perimeter.

WHOSchools · Business · Organisations

Who it’s for

Built for places that hold sensitive data and still want AI help — starting with schools, then businesses and wider organisations.

01

Schools

Pupil data, safeguarding notes and staff AI use under UK GDPR and Department for Education generative AI expectations — without pasting raw records into public chat tools.

AUD-SCHOOLAudience
02

Businesses

HR files, customer records and commercial documents for SMEs that want AI help with policies that keep sensitive text on-premises by default.

AUD-BUSINEAudience
03

Organisations

Charities, councils, healthcare, legal and finance teams that handle regulated or confidential data and need controlled egress, not open internet AI.

AUD-ORGANIAudience

Privacy by appliance, not by promise

Sensitive documents are handled on the box. Cloud AI is optional, airlocked, and limited to allowlisted endpoints.

On-premises by design

The appliance lives in your building. Sensitive source documents do not have to leave the premises for every AI interaction.

Desensitise before cloud

The airlock strips or replaces sensitive fields before an optional cloud model sees a prompt, then rehydrates the answer locally.

Fully local option

When policy requires it, Budbox can run local models on the box GPU so the prompt never leaves the network.

Allowlisted egress only

Outbound AI traffic is limited to endpoints you approve. Unlisted hosts are not reachable from the stack.

What the stack includes

  • Locked-down AI stack in Docker on a dedicated on-premises box
  • Document airlock: desensitise before cloud AI, rehydrate locally
  • Option to run fully local models on a ~12GB NVIDIA GPU-class card
  • Local graph database for domain knowledge enrichment
  • Dexl workflow node, including conditional question steps
  • Egress locked to allowlisted AI endpoints only

Hardware

Dedicated on-premises box with a ~12GB NVIDIA GPU-class card. Typical hardware is from around £1,000–1,500. Configuration depends on workload.

Billing

Billing is hybrid: a hardware lease plus AI credits. Pricing is from / contact us — we size credits and lease to the organisation, not a one-size public plan.

Next step

Enquire with your organisation type and data constraints, or book a short demo chat. We will not invent a public plan on this page.

AI your staff can use without losing the perimeter.

  • Schools: pupil data and safeguarding stay controlled
  • Businesses: HR and customer files stay on the box by default
  • Organisations: allowlisted egress only
  • Dexl workflows with conditional question steps

Budbox, plainly

What is Budbox?

Budbox is a local privacy appliance from OnDraft: a dedicated machine that runs a locked-down AI stack in Docker on your premises. It is for schools, businesses and organisations that need AI without treating sensitive documents as ordinary cloud chat input.

Does everything stay offline?

You can run fully local models on the box. When you choose a cloud model, sensitive documents are desensitised on the box first (the airlock), and only allowlisted AI endpoints are reachable. Raw source files do not need to be the outbound payload.

How does it fit Dexl?

Budbox integrates with Dexl workflows as a special node type, including conditional question steps used in admin and process flows.

What does the hardware cost?

Hardware is typically from around £1,000–1,500 for a dedicated box with a ~12GB NVIDIA GPU-class card. Billing is hybrid: hardware lease plus AI credits. Contact us for a sized quote rather than a public price list.

Who is it for?

Schools handling pupil and safeguarding data; businesses with HR, customer or commercial files; and organisations such as charities, councils, healthcare and legal teams that need UK data-protection-aware AI use.

We build systems that power growth.

AI agents, integrations and operational software—built around your business.