Service 05

Deploy useful AI without surrendering your data boundary.

Inference inside your perimeter: data sovereignty, predictable cost and full control — so the projects legal keeps blocking become the ones you can actually approve.

your perimeter · nothing egressestenancy: yours

INGEST

Your documents

Indexed in place, source permissions carried through.

REDACT

Before inference

Named fields stripped and tokenised, not merely masked in the UI.

SERVE

Open weights

vLLM or Ollama behind an OpenAI-compatible API.

RECORD

Audit trail

Who asked what, over which source, retained to policy.

weights · yourslogs · yoursembeddings · yoursegress · none by default

What we fix

Six reasons AI projects stall at legal.

SOVEREIGNTYData leaves the perimeterCustomer records, contracts and case notes posted to a third party. Legal says no, so the useful projects never start.
COMPLIANCEResidency obligationsContracts or regulators require data to stay in a jurisdiction. A shared API endpoint cannot give you that guarantee in writing.
COSTUnpredictable spendPer-token billing that scales with success. Nobody can forecast next quarter, and the finance conversation blocks the rollout.
CONTROLVendor lock-inPrompts, tools and evaluation tied to one provider. Their deprecation notice becomes your unplanned migration.
RISKShadow AIStaff pasting internal documents into consumer chatbots because no sanctioned tool exists. Banning it does not stop it.
EVIDENCENo audit trailWhen an auditor asks who asked what, over which document, on what date, the honest answer is that nobody recorded it.

Prove it on your data before you buy hardware.

  1. 01

    Feasibility & sizing

    Your obligations, your use cases, your volumes — turned into a model shortlist, a capacity estimate and an honest cost comparison against hosted APIs.

    deliverable: sizing & cost model
  2. 02

    Pilot inside the boundary

    One real workload deployed in your environment, scored against a case set built from your own questions, with redaction and audit logging on from the first day.

    deliverable: scored pilot
  3. 03

    Production & handover

    Infrastructure-as-code, autoscaling or fixed capacity, dashboards, alerts and runbooks — then your team operates it, with us on call only as long as that helps.

    deliverable: operated by you

What a private deployment includes.

Everything lives in your accounts, under your access control, in code you keep.

Deployment optionsYour VPC, a dedicated tenancy or on-premise hardware — chosen against your obligations, not our preference.VPC / on-prem
Open-weight modelsLlama, Qwen, Mistral and similar, sized to the task, served behind an OpenAI-compatible API so application code stays portable.portable interface
Retrieval over your corpusYour documents indexed inside the boundary, with permissions honoured at query time so answers respect who is asking.permission-aware RAG
Redaction & retentionSensitive fields stripped before inference, retention windows enforced, and prompt logs treated as records with a policy.policy-enforced
Evaluation harnessA graded case set from your real questions, run before every model or prompt change, so upgrades are decisions and not surprises.regression-gated
Capacity & cost modelSizing from your real concurrency, GPU or CPU cost modelled per month, with the break-even against API pricing shown honestly.fixed monthly envelope

A self-hosted pricing model serving 180,000+ listings for about $10 a month in infrastructure.

Read the Carmark case study

Private AI questions

For summarisation, extraction, classification, search and most internal assistants, current open-weight models are comfortably good enough. For frontier reasoning they are not equal to the best hosted models. We test that on your own cases during the pilot and tell you plainly which parts belong on which model — a hybrid split, where only non-sensitive work leaves the perimeter, is often the right answer.

Keep the data. Still get the AI.