Skip to content
Last Furlong

Private AI integration

We handle the last mile of AI.

Your model isn't the problem. Getting it inside your network, wired to your systems, and past your security review is.

Model choice, prompt design, benchmark scores. The part that demos well.

Deployment, integration, security review. The part that ships.

AI pilots rarely fail on model quality.

They stall at the boundary. By the time a pilot is ready to become something people depend on, the questions stop being about accuracy and start being about where the data goes.

  • The data can't leave

    If your records can't be sent to a third-party API, most vendor reference architectures stop being available to you on page one.

  • The systems are inside

    The work worth automating lives behind a firewall, on hosts that accept no inbound connections and won't be getting any.

  • Someone signs it off

    A reviewer will ask where data goes, who can decrypt it, and what gets recorded. Vague answers end the project.

What we do

  • Run the model where the data already is

    Local inference through Ollama or vLLM, so prompts and documents stay on your hardware. Hosted APIs remain available where your policy allows them — that's your decision, not a constraint of ours.

  • Reach internal systems without opening ports

    Outbound-only tunnels and reverse SOCKS give the agent a route to private services. Nothing inbound needs to be exposed, so there is no new firewall rule to argue for.

  • Put it where the work happens

    Command line, desktop, mobile, and an end-to-end encrypted chat surface. Tool access to your own systems over the Model Context Protocol.

Built on our own stack

We don't resell a hosted wrapper around someone else's API. Delivery runs on Kavilo — software we write and maintain ourselves, which is why we can change how it deploys to fit your constraints instead of asking you to fit a product.

Single binary
Statically compiled, with nothing else to install or run. Copies onto a host that has no internet access.
Providers
One interface over local inference (Ollama, vLLM) and hosted APIs (Anthropic, OpenAI, Gemini), with retry and circuit breaking between them.
Tools
Filesystem, shell, web and retrieval tools, plus Model Context Protocol clients over stdio and streamable HTTP for reaching your services.
Messaging
A hub whose server never holds plaintext: X3DH and PQXDH key agreement, Double Ratchet, sender keys for groups, and sealed sender.
Networking
Outbound-only HTTP tunnelling, and bidirectional SOCKS5 over TLS or QUIC.
Source
Closed. Third-party inspection and source escrow can both be arranged on request, as priced options.

How engagements start

Three stages, and you can stop after any of them. Most clients begin with Assess, because it's the cheapest way to find out whether we're any good.

Assess

We map where your data can and cannot go, which systems hold the value, and what your reviewers will require. Fixed scope, fixed fee. It ends with a written architecture you own, whether or not you continue with us.

Deploy

We build it: topology of your choosing, connections to your internal systems, a surface your people actually use, and runbooks at handover.

Operate

Upgrades, model and provider changes, and acceptance tests re-run on every change so behaviour is measured rather than assumed.

What each stage includes →

Finish

Start with the boundary, not the model.

An Assess engagement is fixed scope and fixed fee, and ends with a written architecture you own — whether or not you continue with us.