> Bron: https://neuralex.nl/en/showcase/sentinel
> An AI NOC in a box: one dashboard, one log and one AI brain that watches over and diagnoses servers — but never acts outside hard guardrails. In production on our own lab.

Showcase · Sentinel

# An AI administrator that knows when not to act

Monitoring tools measure. Sentinel **reasons and acts** — with a brain that cross-checks signals before anything happens, and guardrails that can refuse any action. In production on our own lab: 94 containers, 210 workflows, one log.

[Request a demo](/en/contact) [See the incident replay](#replay)

4

layers, strictly separated

1

append-only log

210

workflows orchestrated

0

mutations without guardrails

Architecture

## Measuring ≠ deciding ≠ acting ≠ remembering

Four layers, strictly separated. The eyes may never mutate, the brain may only propose, the hands may only act within guardrails — and everything lands in one log.

Incident replay

## Same incident, two outcomes

A real (anonymized) incident from our own lab: a heavy data-processing job let a health check time out. Left: what classic tooling did. Right: what Sentinel does.

### Without a brain

1.  Health check times out under heavy, legitimate load
2.  Strike 1 → 2 → 3: a healthy container gets restarted
3.  Escalation: AI diagnosis looks in the wrong place
4.  Another diagnosis, another restart — costs keep climbing
5.  Root cause stays undiscovered: the context lived nowhere

### With Sentinel

1.  Same signals arrive in the event store
2.  Brain cross-checks: "heavy job active, load is CPU-bound"
3.  Restart refused by the load guardrail
4.  Escalation suppressed — no wasted AI calls
5.  One line in the log, with the full reasoning

Tech in detail

## Trust is an architecture choice

### Guardrails before anything else

The only component allowed to mutate sits behind a whitelist, a deny-list, backup-before-every-action, a cooldown and load awareness. Under high load or in maintenance mode, every action is refused and logged instead of executed.

### Bring your own brain, fully local too

The reasoning brain is model-agnostic: plug in any API, an EU-hosted model, or a model that runs entirely on your own hardware. The brain can only propose — it can never act on its own.

### Human in the loop

Every proposal from the brain becomes an approval card on the dashboard. Only after approval does the bridge execute — within the guardrails. Signal lights make the status visible around the house.

### One incident, one thread

Every signal, every piece of reasoning and every (refused) action gets a correlation ID. You read an incident back as a single story — and the chat answers "why did X restart yesterday?" with the actual chain of events.

### Prompt governance

The system instructions for the AI administrator are visible and editable on the dashboard — with mandatory backup and an audit trail. Your AI has no hidden instructions: you read them, you adjust them.

### Crash-resistant design

The core runs natively on the host system, deliberately outside the Docker stack it watches over. If the container layer goes down, the eyes, the log and the hands keep working.

Built with Event store (append-only)Systemd native servicesn8nModel-agnostic LLM layerApproval flowHue signalling

Early access

## Want to see this running on your own servers?

Sentinel runs in production on our own lab and is coming to homelabs, SMEs and IT admins. Request a demo and see the incident replay live.

[Request a demo](/en/contact) [Our approach](/en/aanpak)

---
Volledige (opgemaakte) versie: https://neuralex.nl/en/showcase/sentinel