LOCAL AI

Choose your sovereignty. From cloud-assisted to fully air-gapped.

Full Brain functionality at the level of network openness you choose. For defense, pharmaceutical, critical infrastructure, and classified facilities. Data sovereignty by design, not by promise.

Not sure which sovereignty level fits?

  • Most commercial deployments → Level 1 (BYOK cloud).
  • GDPR-regulated, NIS2 in scope, internal IP concerns → Level 2 (Local-first).
  • Pharma GMP isolated, defense classified, ITAR-controlled → Level 3 (Fully air-gapped).

Air-gapped • EU AI Act-aligned • Five-tier safety authorization • Full audit trail

SOVEREIGNTY LEVELS

Pick your sovereignty level.

Level 1

Cloud-enabled

Brain calls Anthropic, OpenAI, or Azure with your BYOK keys. Best AI quality, fastest setup. Data leaves your cabinet to your chosen provider; never to Interkey.

Level 2

Local-first

Smart routing. Simple queries (read tags, check alarms, quick lookups) stay on the Jetson. Complex system generation falls through to your BYOK cloud provider only if enabled. Best balance of capability and privacy.

Level 3

Fully air-gapped

No internet. Local LLM via Ollama, vLLM, LM Studio, or any OpenAI-compatible endpoint. No telemetry. No update pings. No phone home. Required for ITAR, TEMPEST, GMP-isolated, and classified deployments.

Brain AI+ subscription

Brain AI+ subscription provides managed cloud AI access. Brain AI+ is incompatible with Level 3 (air-gapped) — air-gapped deployments must use BYOK or fully local models. This is by design. See /pricing.

THE LOCAL STACK

How every AI function runs on your hardware.

04Smart Routing

Simple queries

Local LLM

Read tags, check alarms, quick lookups. Fast, free, private.

Complex generation

Cloud if enabled (optional, BYOK)

Large system generation falls through to the best available provider — only if you allow it.

Routing today

Smart routing today is binary: simple queries local, complex generation cloud (if enabled). Configurable per query type — see settings. Q4 2026: nuanced routing per query class.

03Brain AI Agent

Orchestrator

100+ tools · 5-tier safety · multi-provider routing

02Local AI Runtime

Vision

Brain Vision

OpenCV + custom CV pipelines. Fully air-gapped.

Language

Local LLM endpoint

Ollama, vLLM, LM Studio, or any OpenAI-compatible server.

01Hardware

AI co-processor

NVIDIA Jetson Orin Nano Super

40 TOPS AI inference at the edge.

Compute module

CompuLab SBC-IOT-iMX8Plus

Brain OS runtime. Industrial-grade SBC.

COMPLIANCE MAPPING

Sovereignty level to compliance.

Sovereignty levelWhat it enablesRequired for
Level 1 — Cloud-enabledBYOK direct to providerMost non-regulated commercial deployments
Level 2 — Local-firstSmart routing, sensitive data stays localDSGVO Art. 32, NIS2, internal IP protection
Level 3 — Fully air-gappedNo external network, all inference localITAR, TEMPEST, 21 CFR Part 11 isolated, classified, EU AI Act high-risk

VISION PIPELINE

From bootstrap to autonomous in 8 hours. Zero ongoing cloud dependency.

PHASE 016h

Learning

API-assisted parameter optimization. Uses cloud vision if available. Can be fully local if not.

PHASE 022h

Validation

Cross-checks detections. Tunes thresholds. Ready for autonomous.

PHASE 03

Autonomous

OpenCV + classical CV runs locally on Jetson. Zero API calls. Fully air-gapped.

Callout

The bootstrap phase is optional. You can skip it entirely and train fully locally. Choose your sovereignty level.

Footnote: fully local training works best for high-contrast detection roles. Complex anomaly detection benefits from API-assisted bootstrapping.

PERFORMANCE

Local vs cloud, side by side.

TaskLocal LLM (Llama 3.3 70B on Jetson)Cloud (Claude Opus 4.7)
Read tag value<1s~1s
Generate ST program30-60s5-15s
Diagnose alarm with manual lookup10-20s3-5s
Vision detection<50ms (Jetson)N/A — always local

NO TELEMETRY

Brain sends nothing home. Ever.

No usage analytics. No error reports. No model feedback. No phone-home checks. No “anonymous” usage data. Your cabinet is your cabinet. What happens on Brain stays on Brain.

What Brain does NOT send

  • Usage statistics
  • Error logs
  • Performance metrics
  • Model interactions
  • PLC program contents
  • Sensor data
  • Customer identifiers
  • License activation pings

BRING YOUR OWN KEYS

If you use cloud models, your keys stay on your cabinet.

Brain never proxies AI calls through our servers. Your Anthropic, OpenAI, or Azure key is stored encrypted on the cabinet. API calls go directly from your Brain to the provider. We don’t see them. We don’t log them. We don’t bill them.

How it works

On premise

Your Brain Cabinet

Direct

Provider API

Anthropic · OpenAI · Azure

Response

Back to your cabinet

Not this

Your BrainInterkey ServerAnthropic

Interkey servers are NEVER in the path.

UPDATE STRATEGY

Updates without phone-home.

Default

Brain checks for updates over HTTPS once per week. Updates can be installed automatically or queued for manual approval.

Disabled

No update checks. You manage updates manually via downloaded packages.

Private update server

Run an on-premise update server. Brain pulls from your server only. We provide the server software at no cost for Brain Compliance subscribers.

Update checks contain no telemetry. Just a version query.

SUPPORTED LOCAL MODELS

Run any OpenAI-compatible endpoint.

Ollama

Easiest local deployment. Llama 3.3 70B, Qwen 2.5, etc. Single binary install.

vLLM

Production-grade throughput. GPU-accelerated. OpenAI-compatible API.

LM Studio

GUI for local model management. Good for smaller deployments.

llama.cpp server

Lightweight, CPU-friendly. Runs on the Jetson itself for small models.

Self-hosted GPU server

Dedicated inference box. Highest performance.

Sizing your local inference.

Inside the cabinet

Jetson Orin Nano Super

Up to ~3B parameter models comfortably. Llama 3.2 3B, Phi-3.5, Qwen 2.5 3B. Good for tool calls and tag reasoning.

Adjacent server

Single RTX 4090

Up to ~70B parameter quantized. Llama 3.3 70B, Qwen 2.5 72B. Production-grade for complex generation.

Adjacent server

Multi-GPU

Frontier-class local models. For customers who want cloud-equivalent quality fully on-premise.

Routing note

Brain routes by query complexity. Simple tool calls (read tags, check alarms) stay local. Complex system generation goes to the best available model (local or cloud, if you allow it). See /safety for the five-tier authorization model and /reliability for degraded-mode behavior.

VERIFY YOURSELF

Don’t take our word for it.

  • Run Wireshark on your cabinet's network port. We'll send you a packet capture template showing what should and should not appear.
  • Disconnect the WAN cable for 24 hours. Brain continues operating. Verify yourself.
  • Inspect the Brain OS network configuration. We provide systemd unit files and nftables rules for review.
  • Audit our update server endpoints. URLs are documented and stable.
  • Run an air-gap deployment trial — request the deployment guide.

WHO BUYS LOCAL AI

Regulated industries. Classified environments. Sovereign operators.

Defense

Classified facilities. ITAR compliance. TEMPEST environments. Brain runs with no network, no telemetry, no cloud dependency. TEMPEST-shielded cabinet variant available for classified deployments — contact compliance@interkey.com.

Pharmaceutical

21 CFR Part 11. GMP validation. Data integrity. Brain's audit trail and air-gap support qualify for regulated production.

Critical Infrastructure

Water treatment. Power grid. Chemical processing. Brain operates isolated from IT networks by design.

Sovereign Industrial

Nation-state manufacturing. Local data laws. Export controls. Run Brain entirely within your borders.

COMPLIANCE

Data sovereignty is not marketing. It’s architecture.

  • All data stays on the cabinet by default
  • No hidden network calls (audit with Wireshark — we'll help you)
  • On-premise update servers supported
  • Air-gap mode is a runtime flag, not a product tier

Your data. Your cabinet. Your control.

READ ABOUT SAFETY