Air-gapped AI, deployed behind your walls.

We design, ship and hand over private models, retrieval and agents that never call out.

transfer-manifest.yaml
Outbound connections: blocked
Connected staging replica
  • model-weights/Staged
    Open-weight LLM, quantized for your GPUs
  • images/Staged
    Inference server and vector database
  • index/Staged
    Embeddings of your internal documents
  • agents/Staged
    Agent runtime and tool definitions
The gap
One-way, approved transfer
Inside your network
  • model-weights/Waiting
    Awaiting transfer
  • images/Waiting
    Awaiting transfer
  • index/Waiting
    Awaiting transfer
  • agents/Waiting
    Awaiting transfer

What is air-gapped AI?

Air-gapped AI is artificial intelligence that runs on hardware inside a network with no connection to the internet. The models, the software around them and the data they read all stay behind the boundary. Nothing calls a cloud API, sends telemetry or pulls an update on its own.

Teams choose it when the data cannot leave: classified programs, patient records, privileged legal files, trading systems, plant controls. The catch is that everything the AI needs has to be brought in on purpose, and that is where most projects get stuck.

Private cloud (VPC)

Where it runs
Provider hardware, in an isolated account
Internet connection
Yes, through controlled routes
How models and updates arrive
Pulled from public registries
Who can reach the data
You, on the provider's infrastructure
Fits when
Data may live in a provider's cloud

On-premises

Where it runs
Your own servers
Internet connection
Usually, for updates and licensing
How models and updates arrive
Pulled or mirrored
Who can reach the data
You and anything your network allows
Fits when
Data must stay on your hardware

Air-gapped

Where it runs
Your own servers, on an isolated network
Internet connection
None
How models and updates arrive
Carried across through an approved transfer
Who can reach the data
Only people and systems inside the boundary
Fits when
Data must never touch an outside network

Why air-gapped AI projects stall

The model is rarely the hard part. Everything around it is.

Dependencies that phone home

Package managers, container registries and model hubs all assume an internet connection. One missing package or base image stops the install.

Weights too big to carry casually

Model files run to many gigabytes and have to be checked, signed and logged at every hop through your transfer process.

GPU drivers that will not line up

Driver, CUDA and inference server versions have to match exactly, and you cannot search for the fix from inside.

No cloud APIs to lean on

Embeddings, evaluation, guardrails and monitoring that normally call a hosted service have to run locally instead.

Updates with no way in

A system that works on day one still needs patched images and newer models, and few teams plan how those cross the gap.

Accreditation paperwork

Security reviewers need every component, where it came from and what it can reach. Missing records hold up approval.

What we deliver

An AI system that runs inside your boundary, plus the records and processes to keep it running there.

fit: task × hardware × license

Architecture and model selection

We pick open-weight models and a serving setup that fit your task, languages, hardware and license terms, then size the GPUs to match.

mirror: packages, images, weights

An offline supply chain

We mirror every package, container image and model weight you need, pin versions, and produce a signed, checksummed transfer bundle.

retrieval: internal documents only

Private LLM and RAG on your hardware

Inference server, vector database and document pipeline on your machines, answering from your internal documents with sources cited.

tools: inside the boundary

Agents that work offline

Agents that call only internal tools and systems, built on the approach behind Papaya's own compliance agents.

evals: run locally

Evaluation and monitoring inside

Test sets drawn from your real tasks and run locally, plus logging and monitoring that never leave the network.

runbook: update, verify, roll back

Handover and an update path

Runbooks, a component inventory for your security review, and a documented way to bring in patches and new models.

How an engagement runs

Most of the work happens where we can move fast. Only a tested, verified system crosses into your network.

  1. With your team

    Map the boundary

    We learn your network rules, hardware, transfer process and the jobs the AI has to do, then write a plan your security team can review.

  2. Connected staging

    Build on a replica

    We build and test the whole system on a connected replica of your environment, so problems get solved where we can still look things up.

  3. Across the gap

    Carry it across

    Everything goes into a verified transfer bundle that moves through your approved process, with a log of every file.

  4. Inside your network

    Run it, then hand it over

    We install, test against your real tasks, train your team and leave runbooks for updates and rollbacks.

Built for teams whose data can't leave

Government and defense

Programs that handle classified or controlled information on isolated networks.

Healthcare and life sciences

Clinical notes, patient records and research data that cannot go to an outside model.

Financial services

Trading, risk and customer data held to strict residency and audit rules.

Legal

Privileged case files and discovery material that need search and drafting help without leaving the firm.

Energy and critical infrastructure

Operational networks kept apart from the internet by design.

Research and manufacturing

Designs, formulas and lab data that are the business itself.

Why Papaya

Papaya builds AI agents that audit privacy and accessibility compliance by running real browser sessions. Our enterprise deployments already offer API, MCP, local and web entry points, with storage and the LLM provider chosen by the customer.

Air-gapped work is the same discipline taken all the way: know every component, control every connection, keep the evidence.

Backed by 1616 Ventures, Ritual Capital and Night Capital. Part of the Launch powered by KPMG accelerator.

Air-gapped AI, answered

What is air-gapped AI?

Air-gapped AI is an AI system, such as a large language model, a retrieval pipeline or an agent, that runs entirely on hardware inside a network with no connection to the internet or any outside network. Models, software and data come in through a controlled transfer process, and nothing the system reads or produces leaves the boundary.

How is air-gapped AI different from on-premises or private cloud AI?

Private cloud AI runs on a provider's hardware in an isolated account, and on-premises AI runs on your own servers. Both usually keep some network path for updates, licensing or monitoring. An air-gapped deployment removes that path entirely, so every model, package and update has to cross the gap on purpose.

Can large language models run fully offline?

Yes. Open-weight models run on your own GPUs with no call to an outside API. The real work is choosing a model that fits your hardware and task, packaging it with everything it depends on, and testing it before it goes inside.

Which models can you deploy?

Open-weight models whose licenses allow your use, chosen for your task, languages and hardware. Hosted models that only work through a vendor's API cannot run air-gapped, so we help you find an offline model that does the same job.

What hardware do we need?

It depends on model size, how many people use it at once and how fast answers need to come back. We size the hardware during the assessment and plan around GPUs you already own where we can.

How do updates and new models get in?

Through the same approved transfer process as the first install. We document a repeatable update path: build and test outside, verify checksums and signatures, carry the package across, and roll it out inside with a way to roll back.

Do you need access to our network?

We work within your access rules. Most of the build happens on a connected staging replica of your environment. The final install follows whatever your security policy allows, whether that means your team runs our runbook or we work on site under your supervision.

Can AI agents work in an air-gapped network?

Yes, as long as every tool they call also lives inside. Papaya builds browser-using AI agents for compliance teams, and our enterprise deployments already support local entry points and a configurable LLM provider. We bring the same approach to agents that work with your internal systems.

How long does an air-gapped AI deployment take?

It depends on scope, hardware and how your organization approves transfers. After the assessment you get a written plan with stages, owners and what we need from your team.

Tell us what you need to run offline

Share a little about your environment and what you want the AI to do. An engineer from our team will follow up to set up a scoping call.

Please keep classified, regulated or otherwise sensitive details out of this form. We will set up the right channel for those.

For example: a chat assistant over internal documents for a team of analysts.