Pilots begin late 2026 Pilot pricing

AI that never
phones home.

Open-weights models, retrieval over your internal documents, and assistants wired into the tools you already run, all of it on your own GPUs, behind your own firewall. Not a single token leaves your network. Not one.

Join the pilot list See what's included
Zero egress, air-gappable Open-weights models, yours Pilots from $8,000
Why private AI

Your documents are
the one thing you can't leak.

Contracts, patient records, case files, source code, the documents that make AI useful are exactly the ones you can't paste into a cloud chatbot. Open-weights models closed that gap: frontier-class capability now runs on hardware you own.

  • Zero egress, provable. The inference stack runs with no outbound network access, your security team can verify it at the firewall, not in a vendor's promise.
  • No per-token meter. Once the hardware is yours, a million queries cost the same as ten, electricity.
  • Model-agnostic. Llama, Mistral, Qwen, Gemma, we benchmark on your tasks and swap models as better ones ship. You're never married to one lab.
  • Wired into your tools. Draft replies in Zammad, summarise records in Odoo, answer questions over your Docusaurus KB, the same stack we already deploy.
Aërgap AI · one stack
on your GPUs
LLM serving
RAG search
Doc ingestion
Chat UI
SSO & roles
API
Guardrails
Usage logs
Replaces Cloud APIs Per-token bills DPA anxiety
What we deploy

Not a demo.
A working assistant.

A pilot ends with something your team uses daily, a private assistant answering real questions over real documents, with the security review already done.

Model selection on your tasks

We benchmark open-weights models against your actual documents and workflows, not leaderboards, and size the hardware to match.

RAG with citations

Answers grounded in your documents, with sources linked on every claim. Access control mirrors your existing permissions, people only get answers from files they could open anyway.

Wired into your tools

Ticket-draft suggestions in Zammad, record summaries in Odoo, KB answers from Docusaurus, assistants live where work happens, not in another tab.

Security review, pre-answered

Network diagram, data-flow documentation and an egress test your security team can run themselves. Air-gapped variants for classified environments.

Your GPUs or ours to start

Run on your existing hardware, a private-cloud GPU tenancy in your account, or start on loaner hardware while yours is on order.

Usage you can measure

Query logs, adoption dashboards and quality feedback loops, so after 90 days you know whether it earned its keep, with numbers.

The pilot

Ninety days from idea
to daily habit.

AI projects die in proof-of-concept purgatory. Ours is scoped to end with a working assistant and a go/no-go decision you can defend.

Scope the use case

Pick one workflow where AI can prove value, support drafting, contract review, internal Q&A. Define what "working" means, in numbers.

Week 1 – 2

Deploy the stack

Inference, retrieval, chat UI and SSO on your infrastructure. Your documents ingested, access controls mirrored, egress test passed.

Week 3 – 5

Tune with your team

Real users, real questions, weekly iteration on retrieval quality and prompts. Model swaps if benchmarks say so.

Week 6 – 10

Measure & decide

Adoption and quality numbers on the table. Scale it, extend it to a second workflow, or shut it down, the stack is yours either way.

Week 11 – 13
Honest pricing

One pilot.
One price. No meter.

Pilots from $8,000 fixed, scoping, deployment, tuning and the 90-day measurement, on your hardware or a GPU tenancy in your cloud account. No per-token fees, no per-seat fees, and the entire stack stays yours whatever you decide. Pilots begin late 2026; the list is first-come, first-served.

After the pilot: $50 / hour consulting to extend, or fixed $150 / month managed care (monitoring, upgrades, model refreshes).
Join the pilot list →
FAQ

Questions
security teams ask.

If yours isn't here, a 20-minute call is usually faster than more reading.

Talk to a human →
For the workflows we pilot, document Q&A, drafting, summarisation, extraction, yes, comfortably. Today's open-weights models match what the frontier looked like 12–18 months ago, and grounding them in your documents with retrieval matters more than raw model size. We benchmark on your tasks before committing, and if the numbers aren't there, we tell you.
Less than you'd think. A team assistant over internal documents typically runs on one or two workstation-class GPUs, a single server, not a cluster. We size it during scoping, and you can start on a private GPU tenancy in your cloud account (or our loaner hardware) while your own is on order.
At your own firewall. The stack runs with no outbound network access, model weights, retrieval index and logs all live on your storage. We hand your security team a network diagram and an egress test they run themselves. For classified environments, the same stack installs fully air-gapped from physical media.
We're hardening the stack with design partners now. Join the pilot list and we'll scope your use case in advance, benchmark candidate models on your documents, and reserve you a slot in the first pilot cohort. Scoping calls are free either way.

The name is not a metaphor. Neither is the firewall.

Free 30-minute scoping call. You leave with a written pilot plan, a hardware estimate and a fixed quote, even if you decide the cloud API is fine for now.

Book a scoping call Join the pilot list →