ArchitectingDigitalEmpires

Jay Wildes — Systems

02 / THE LAB

Under one roof

Enterprise hardware, owned and run end to end. The fabric carries a hundred gigabits to the storage and model servers, the edge inspects everything that leaves, and the whole of it is cooled by a loop I fabricated.

100 Gb to storage and model servers measured 2026-08-09

10 Gbps inspected at the edge measured 2026-08-09

6 kW+ closed-loop cooling capacity measured 2026-08-09

A fully populated 42U enterprise rack photographed square-on in a dark room: cable management and a patch panel at the top, a switch, a stack of thin compute nodes, two large drive chassis, and a power unit at the base.

42U · front elevation · scroll the rack

  1. 01 / 06

    CORE

    Switching fabric. Everything in the rack meets here, and nothing crosses between segments without being told to.

    running
    Layer 3 switching, segmented per trust level
    rate
    100 Gb to storage and the model hosts · 10 Gb to everything else

  2. 02 / 06

    EDGE

    The only path in or out. Nothing reaches the internet without passing through it first.

    running
    Custom NGFW and router, nftables, deep packet inspection, and a parser that learned what normal traffic looks like
    rate
    10 Gbps, fully inspected

  3. 03 / 06

    CLUSTER

    The control plane. Immutable by design, so there is no shell to log into and no state to drift.

    running
    Talos Linux, Kubernetes. UMBRA rides on this.
    rate
    10 Gb

  4. 04 / 06

    MODELS

    Local inference. Nothing leaves the building to answer a question, which is the whole reason the rest of this exists.

    running
    Local models, the assistant's reasoning, the voice agent's grounding
    rate
    100 Gb to storage

  5. 05 / 06

    STORAGE

    Datasets, model weights, and every log the parser reads. Sized so the model hosts never wait on it.

    running
    Bulk storage and snapshots
    rate
    100 Gb

  6. 06 / 06

    POWER · THERMAL

    Conditioning and ride-through at the base. The heat everything above it makes leaves through the immersion loop rather than into the room.

    running
    UPS, and the closed cooling loop
    rate
    over 6 kW of loop capacity
10 Gb inspected100 Gb100 Gb10 Gbthermal

select a node

  • WAN — Upstream. Nothing trusted arrives here.
  • EDGE — Custom NGFW and router. Deep packet inspection at ten gigabits. The logs are parsed by something that learned what normal looks like.
  • CORE — Switch fabric. Hundred-gigabit links to the machines that need them, ten to everything else.
  • MODELS — Local model servers. Submerged. A hundred gigabits to storage, because a model that waits on disk is a model you are not using.
  • STORE — Storage servers. Submerged, and on the same hundred-gigabit fabric as the model hosts.
  • TALOS — Kubernetes on immutable infrastructure. No shell to log into. This is what UMBRA runs on.
  • ASSISTANT — Voice on site, and remotely through an Android app I wrote. Persistent context. It rewrites its own tooling.
  • VOICE — Answers the phone. Grounded against a knowledge base so it declines to invent an answer.
  • TANK — Single-phase immersion, closed loop, fabricated here. Over six kilowatts of capacity. Nothing in it spins.

automation

  1. 01 ingest wearable, blood panel, firewall log, transaction
  2. 02 normalise one schema, timestamped, versioned
  3. 03 reason local models, on submerged hardware
  4. 04 act plans, blocks, calls, messages
  5. 05 review what it did, and what it changed about itself

03 / WORK

Five systems

01 / UMBRA

A VPN service I designed and run. VLESS with XTLS Vision, on Talos Linux because the nodes should not have a shell to log into. The control plane is provider-agnostic, so a VPS company losing interest in me is a scheduling problem rather than an outage.

UMBRA · VLESS/XTLS Vision · Talos Linux · immutable · provider-agnostic

02 / EDGE

A next-generation firewall and router I built for the network edge. Deep packet inspection at line rate. I read the logs for a year, and then something else started reading them.

10 Gbps inspected throughput measured 2026-08-09

EDGE · deep packet inspection · nftables · log parsing

03 / IMMERSION

Single-phase immersion cooling, designed and fabricated. A closed loop that takes heat out through the liquid instead of the air. Larger systems are in build.

6 kW+ loop capacity measured 2026-08-09

IMMERSION · single-phase · closed loop · no fans in loop

04 / VOICE

An agent that answers the phone and holds the conversation. Grounded against a knowledge base so it declines to invent an answer.

VOICE · real-time · knowledge-base grounded

05 / ASSISTANT

Voice-accessible on site and remotely through an Android app I wrote. It holds persistent context about my life, sees through the cameras, and rewrites its own tooling. It has context I never gave it.

ASSISTANT · persistent context · vision · self-improving

Eight server boards suspended in a glass bath of clear dielectric fluid, silhouetted against a blown-out backlight. Every board is passively cooled with bare finned heatsinks. The fluid surface is unbroken.

04 / IMMERSION

The heat is gone

Boards sit in dielectric fluid. The heat leaves through the liquid instead of the air, so there is no fan anywhere in the loop and nothing to hear. Larger systems are in build.

single-phase · dielectric · closed loop · silent

Before any of this

I started by fixing other people's machines. PrimePC was a repair and consulting business I owned and ran in Lexington, South Carolina.

You learn what fails. Not what should fail, or what the datasheet says fails. What actually comes across the bench, over and over, until you can tell from the sound it makes.

05 / ABOUT

Jay Wildes in a dark collared jacket, standing in a dim server hall and lit only by the blown-out light at the far end of it.

One operator

There is no team. The scale here is normally staffed and it isn't, and the reason is a set of AI workflows I built and tuned for my own use. Not an assistant I subscribe to. Tooling I own, running on hardware I own.

That is the same problem an AI company is hiring against, which is most of why this site exists.

one operator · own tooling · own hardware