Case study · Network & systems build-out

The systems I build,
and run every day.

One person, a small server room, and a fleet of AI agents. A virtualized server, a segmented network, a monitoring appliance, a GPU node for local AI, and nightly backups that are read back and checked, all running a live market-data platform around the clock.

Anonymized on purpose: machines are shown by role. No names, addresses or accounts.

~10 minbetween backups of the business documents
24/7machines, services and the internet watched around the clock
Nightlyarchives read back and hash-checked; restores tested on purpose
3,600+commits across the platform in ten months (as of Sept 2026)

How it fits together

data telemetry backups
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 · 48 GB market data automation development windows GPU inference node 3 GPUS · LOCAL LLM · SPEECH · RENDER SENTINEL OBSERVABILITY · ~20 CONTAINERS Backup NAS RAID1 · RECOVERY SNAPSHOTS Workstation VOICE-DRIVEN LINUX Media rooms VOICE-CONTROLLED

Swipe sideways to see the whole map →

01 · Foundation

Network & servers

Built like a small business, not a hobby shelf. One tower server runs four virtual machines, each with a single job. The network is split into planned address bands, with an isolated segment for untrusted devices checked against a written isolation report.

Virtualization

  • One host, four VMs: market data, automation, development, Windows
  • Fixed memory per VM, RAID storage
  • VM state, checkpoints and disk sizes read every minute

Segmented network

  • Zone-based firewall with intrusion detection
  • Planned address bands per role
  • Isolated VLAN, isolation tested and documented

Names & certificates

  • Internal services on trusted TLS certificates
  • Internal services are reachable only from the local network
  • Only what must be public goes out, through secure tunnels
02 · The platform

A live market-data platform

The busiest machine in the room. It serves real-time quotes, options analytics and a research platform to dashboards and AI agents, around the clock.

APIs

  • Quotes, options chains, volatility and options-flow analytics
  • Market regime, research summaries and per-ticker intelligence
  • A price-alert registry checked every 15 minutes, pushed to my phone

Always running

  • Every service supervised; scheduled jobs pinned to New York market time
  • A paper-trading bot and a live charting app
  • Every database backed up nightly

Automation server

  • Supervised scanner services with circuit breakers
  • Options-flow capture every 15 minutes during market hours
  • One notification hub for alerts, email and calendar
03 · Observability

SENTINEL: it explains, then stays quiet

A low-power appliance watches the machines, services and the internet connection. The rule it is built on: 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.

How SENTINEL listens

signalscollecteddecided
WHAT IT WATCHESHOW IT LISTENSWHAT IT DECIDESWHAT YOU GET Servers & VMsCPU · MEMORY · DISKSNetwork & Wi-FiGATEWAY · SWITCH · APSServicesIS IT UP? IS IT RIGHT?The internetIS THE LINK HEALTHY?Logs & changesLOGINS · ACCOUNTS · FILESMetricsNUMBERS OVER TIMEFlows & SNMPWHO TALKS TO WHOMHealth probesFROM INSIDEPublic-name probesBY PUBLIC ADDRESSLog collectionSEARCHABLE HISTORY SENTINEL ALERTS AROUND THE CLOCK NOISE → ONE INCIDENT EVIDENCE ATTACHED explains, then stays quiet One clear alertWITH THE EVIDENCE · TO MY PHONE DashboardsEVERYTHING AT A GLANCE SilenceWHEN ALL IS WELL WATCHDOGS ON 2 OTHER MACHINES

Swipe sideways to see the whole map →

The stack

  • Metrics, logs, alerts and dashboards in one appliance
  • Network flow, SNMP, disk health and service probes
  • Alert rules for every machine, service and the internet link

Incidents, not noise

  • Raw alerts become one incident, with the evidence attached
  • A dead-man check on two other machines watches the watcher
  • The dead-man alarm passed a 19-case test before it went live

Security checks

  • New devices, failed logins and account changes
  • File integrity, pending updates, newly opened ports
  • Replaced a heavyweight security suite with 6 lean checks
04 · Recovery

Backups you can actually restore

A backup you haven't restored is a hope, not a backup. The nightly archives are read back and hash-checked, and restores are tested on purpose.

Every night

  • Per-machine archives and live database copies to the NAS
  • Each one read back and hash-verified
  • Kept 14 days, 8 weeks and 12 months

Versioned & recoverable

  • 30-day recovery snapshots
  • Weekly versioned backup of the whole NAS
  • Each machine's backup key works only from that machine

Proven restores

  • Business documents backed up about every 10 minutes
  • A test restore came back byte-identical (Oct 2026)
  • One-file restore tool and a restore-test script
05 · AI operations

AI agents run it, safely

Most of this is operated through AI agents, by voice. That only works with guardrails, so every agent works inside them: look first, change carefully, write it down, and keep a way back.

Guardrails

  • Agents inspect machines read-only first
  • Every change logged with how to reverse it
  • Snapshots and backups before risky work

Least privilege

  • Read-only keys where reading is enough
  • Keys locked to the machine that uses them
  • VM control designed as a short allow-list with a human approval step

Local AI

  • Language models and speech-to-text on my own GPUs
  • A job queue that shares the GPUs fairly
  • A live view of Claude Code sessions on five machines (built on StrSimon's AgentVille)
"I know where AI gets IT work right and where it breaks, because I run it on real systems every day."
Daryll Gomas · 25 years in IT