Nerve Centre
Internal R&D PlatformOur sovereign AI operations platform.
Nerve Centre is a self-hosted, modular AI operations platform that runs on company-owned hardware with a local GPU. It acts as our internal AI-ERP, orchestrating agent workflows, measuring every inference call, and strictly isolating workloads from the host operating system.
How it operates
A Nuxt 3 control plane shows every model call, what it cost, and where it ran. All models are reached through a central gateway that enforces per-caller keys, budgets, and rate limits. The system prefers local GPU inference (e.g. Qwen 2.5 Coder 7B, Llama 3) and only routes work to cloud models by explicit policy.
AI agents act as executors. They propose changes as source-controlled pull requests. An AI reviewer evaluates the scripts, and a human approves every privileged step with a hardware security key before a central orchestration service runs the workflow.
Architecture Principles
Sovereign by Default
Local GPU inference is preferred via a dedicated LiteLLM gateway. Paid cloud usage requires explicit policies, and data classification ensures only public/code data hits free cloud tiers.
Human-in-the-Loop
A human approves every privileged step with a YubiKey hardware key assertion. AI execution agents propose changes via pull requests, but never execute privileged changes.
Everything Measured
Every model call is recorded (caller, model, tier, tokens, averted cost). Prompts and answers are deliberately not stored. We track our Sovereignty ratio live.
Containment
Runs inside a strictly unprivileged LXD container with nested Docker. All egress is default-deny, forcing traffic through a dedicated allowlisting proxy.
We built this to run our own company. Now we're making the foundation available.
Explore Sandbox As A Service