SERVICE 02 · OPERATE
AI that runs on your machines, on your terms, on the record.
ARTVS is an execution control plane for AI document work. It routes each task to the least expensive model that handles it reliably, keeps a per-call ledger of what everything cost and how it turned out, and can run entirely inside your network.
The core idea
Route by evidence. Account for everything.
Different tasks need different models. Simple classification runs well on small local models that cost nothing per call; harder reasoning earns a frontier model. ARTVS routes each task type to the cheapest tier that our benchmarking shows handles it reliably — and because every routing decision came from measurement, we can show you why each one was made.
Every task writes one row to a ledger: what ran, on which model, what it cost, and how it turned out. The number we manage to is cost per verified success, not cost per token. When you ask what your AI spending bought you, the answer is a query, not an estimate.
Architecture
Built like something that expects to be audited.
Local-first by design
The stack runs on hardware you control. Local models process documents without anything leaving your network; external model calls happen only through a single sanctioned path, and only for task types routed there deliberately.
One way out, and it's watched
Exactly one component of the system has an outbound internet route. Everything else is boxed in at the network level. The audit ledger lives on separate storage that the internet-facing component cannot touch.
Routing profiles, not lock-in
Model tiers live in one configuration file. Swapping providers, adding a new model, or restricting to US-based providers only is a profile change, not a rebuild. A fully local profile is available for work that must never leave the building.
Version-pinned components, backup-before-update, documented rollback, and a daily health signal come standard. Security documentation is written against NIST 800-53 control families; the security narrative and model provenance documentation are available on request.
What deployment looks like
Scoped to your work, proven before it's trusted.
Deployment starts from assessment findings: the task types worth automating, in your volumes, with your definitions of correct. The system is configured for those task types, validated against test documents before it touches real work, and handed over with the ledger running from day one.
You see the same numbers we do. If a task type underperforms, the ledger shows it, and the routing changes or the task comes back out of automation. The system is designed to earn trust task type by task type, not to demand it up front.
Pricing
Setup once, run it yourself, call us when you need us.
Deployment Setup
Hardware spec, stack installation, model selection, routing configuration, ledger setup, and validation against your task types. Handed over running.
Starting at $4,500Managed Operation
Monthly health checks, model updates, routing adjustments from ledger data, and priority support. Your system stays current without your team maintaining it.
Starting at $600 / monthDeployments are scoped from assessment findings. Setup pricing scales with the number of task types and whether hardware procurement is included. The managed tier is optional — the system is designed to run independently once deployed.
Contact
Talk through a deployment.
Deployments are scoped from assessment findings, so the conversation usually starts there. If you already know your task types and volumes, we can start from those instead.
Email ARTVS