Daryll Gomas
Running · private

SENTINEL

Monitoring that explains, then stays quiet.

A low-power appliance that watches the machines and services, and the internet link, in a small server room. It turns raw alerts into one incident with the evidence attached, and tells me only the few things I'd want to know.

MonitoringSecurity detectionPrometheusLokiGrafanaAlertmanager
Example alert
SENTINEL · incident3m open
CriticalConfidence: high

Virtualization host is down. 4 machines affected.

Virtualization host stopped answering at 10:14 AM. Market-data VM, Automation VM, Development VM and Windows VM depend on it, so they went down with it. The rest of the network is fine.

Market-data VMAutomation VMDevelopment VMWindows VM
Evidence
registry checkNo answer from Virtualization host (address removed).
alertmanagerHostDown
What to doCheck that Virtualization host has power and a network cable, and whether it is in the middle of a reboot.
Example · names replaced by roles
18hosts watched
29services watched
19test cases passed by the dead-man alarm
6lean security checks replaced a heavy suite
2other machines watch the watcher
01The problem

More alerts is not more safety.

One person runs a small server room: a virtualization host with four VMs, a segmented network, a GPU node, a backup NAS, and a market-data platform that has to stay up around the clock. All of it needs watching, and so does the internet link.

The usual answer is more detection and a heavier security suite. But a monitor that pages you for everything trains you to ignore it.

Success isn't detecting more things. It's noticing the few things a person would want to know, explaining them correctly, and staying silent the rest of the time.

02What I built

One quiet appliance that watches everything.

SENTINEL is a low-power observability appliance (about 20 containers on the box, with its control plane, as of Oct 2026) that collects metrics, logs, network flows and health probes from every machine and service, then decides what is worth a person's attention.

  • Raw alerts are correlated into one incident with the evidence attached, sent to my phone.
  • Security checks flag new devices, failed or unusual logins, new accounts, file-integrity changes, pending updates and newly opened ports.
  • Public web addresses are probed by their public names, and the internet link itself is measured.
  • A dead-man check on two other machines watches the watcher, so a silent SENTINEL is never mistaken for a healthy one.
03How it works

Signals in, one clear answer out.

How SENTINEL works: machines and services send signals; SENTINEL collects them, turns raw alerts into one incident with evidence, and sends one clear alert, dashboards, or silence. WHAT IT WATCHES HOW IT LISTENS WHAT IT DECIDES WHAT YOU GET Servers & VMs CPU · MEMORY · DISKS Network & Wi-Fi GATEWAY · SWITCH · APS Services IS IT UP? IS IT RIGHT? The internet IS THE LINK HEALTHY? Logs & changes LOGINS · ACCOUNTS · FILES Metrics PROMETHEUS Flows & SNMP WHO TALKS TO WHOM Probes INSIDE + BY PUBLIC NAME Log collection LOKI SENTINEL ALERT RULES · ALERTMANAGER NOISE → ONE INCIDENT EVIDENCE ATTACHED explains, then stays quiet One clear alert EVIDENCE · TO MY PHONE Dashboards GRAFANA Silence WHEN ALL IS WELL Watchdogs on 2 other machines A DEAD-MAN CHECK WATCHES THE WATCHER How SENTINEL works: machines and services send signals; SENTINEL collects them, turns raw alerts into one incident with evidence, and sends one clear alert, dashboards, or silence. WHAT IT WATCHES HOW IT LISTENS WHAT IT DECIDES WHAT YOU GET Servers & VMs CPU · MEMORY · DISKS Network & Wi-Fi GATEWAY · SWITCH · APS Services IS IT UP? IS IT RIGHT? The internet IS THE LINK HEALTHY? Logs & changes LOGINS · ACCOUNTS · FILES Metrics PROMETHEUS Flows & SNMP WHO TALKS TO WHOM Probes INSIDE + BY PUBLIC NAME Log collection LOKI SENTINEL ALERT RULES · ALERTMANAGER NOISE → ONE INCIDENT EVIDENCE ATTACHED explains, then stays quiet One clear alert EVIDENCE · TO MY PHONE Dashboards GRAFANA Silence WHEN ALL IS WELL Watchdogs on 2 other machines A DEAD-MAN CHECK WATCHES THE WATCHER
Animated diagram. Machines are shown by role: no names, addresses or accounts.

Watches

  • Servers and VMs: CPU, memory, disks
  • Gateway, switch and Wi-Fi access points
  • Every service: is it up, and is it right?
  • The internet link
  • Logins, accounts and file changes

Listens

  • Metrics over time (Prometheus)
  • Network flow and SNMP: who talks to whom
  • Disk health and service probes, inside and by public name
  • Searchable log history (Loki)

Decides

  • Alert rules for every machine, service and the internet link
  • Raw alerts become one incident
  • The evidence is attached to it

Delivers

  • One clear alert, with evidence, to my phone
  • Dashboards: everything at a glance (Grafana)
  • Silence, when all is well
04Tools

Standard parts, used plainly.

Prometheusmetrics over time
Alertmanageralert routing and grouping
Lokilog collection and search
Grafanadashboards
Flow export + SNMPnetwork traffic and devices
Health probesinside and by public name
Linux + containersabout 20 on the box, with the control plane
Claude Code, CodexAI writes the code; I specify and verify
05Proof

What it covers today.

18 · 29hosts and services watched around the clock
2other machines run the dead-man check that watches SENTINEL
6lean security checks in place of a heavyweight security suite
Where it sits in the build-out
Internet TLS · SECURE TUNNELS Gateway & firewall ZONE FIREWALL · IPS · FLOW EXPORT 48-port PoE core switch SEGMENTED ADDRESS PLAN · ISOLATED VLAN · WI-FI 7 Virtualization host TOWER SERVER · RAID · 4 VMS market data automation development windows GPU inference node 3 GPUS · LOCAL LLM · SPEECH · RENDER SENTINEL THIS PROJECT · ~20 CONTAINERS Backup NAS RAID1 · RECOVERY SNAPSHOTS Workstation VOICE-DRIVEN LINUX Media rooms VOICE-CONTROLLED
datatelemetry into SENTINELbackupsSwipe sideways, or tap Enlarge

The live dashboards stay private: they show machine names and addresses. This diagram is from the anonymized network case study, shown by role.

06How it's tested and run

Tested, not trusted.

  1. The alarm was tested before it shippedA 19-case scenario test of the dead-man alarm passed before deploy, then an independent AI reviewer checked it twice.
  2. The watcher is watchedA dead-man check on two other machines alerts if SENTINEL itself goes quiet.
  3. Probes by public namePublic web addresses are also checked by their public names, not only by their inside addresses.
  4. Built with AI, operated with guardrailsAI tools write the code; I own the requirements, the verification and the operations. Agents inspect read-only first, and every change is logged with how to reverse it.
07What I'd do next Proposed · not built

Make the config code, then share it.

  • Keep the alert rules, dashboards and host setup in version-controlled configuration, deployed with Ansible: the same 18 hosts, no drift.
  • Publish a cleaned-up copy of the rule set, so others can run the same quiet setup.
08Skill tags
MonitoringSecurity PrometheusAlertmanagerLokiGrafana SNMPLinuxContainersIncident responseAI-assisted build