Theme
Field Manual · Edition 001 · Three Layers, One Owner · Effeciency + Capability + Security

CLASSIFICATION: SOVEREIGN // DISTRIBUTION: YOU DECIDE

Personal
Data Center

Improvise. Adapt. Self-host.

Three layers, one owner. A Start9 server at the hub. Local GPUs for inference and training. Cloud APIs as a tool, not a landlord. Your data stays home. You control how to best interact with AI. Effeciency, Capability, and Security is Your Leverage.

Read the Brief Open OP 01

YOUR HOME NETWORK vLAN (only you) THE OPEN WEB VPS out front selective reveal START-TUNNEL home IP hidden encrypted tunnel START9 SERVER — StartOS OpenClaw Gitea Your Apps Daily Vol THE HUB — everything routes through here YOUR DEVICES PWA — native feel, your box behind it phone · laptop · anywhere LAN LAN LOCAL GPU inference · serving daily cron jobs always on · no API cost LOCAL GPU fine-tuning · training your own weights data never leaves home 200 Gbps interconnect augments when needed CLOUD GPUs on-demand burst capacity API calls · rent as needed

Bayside High · Computer Lab · Est. wherever your server lives

Saved by
the PDCPersonal Data Center

Your compute. Your keys. Your off switch.

Three layers, one owner. A Start9 server at the hub. Local GPUs for the heavy lifting. Cloud APIs when you need the best reasoning, but your data stays home. Build what you need, run it on hardware you own, and reach it anywhere.

See Today's Schedule Enter Homeroom

The Build Log

// four operations, and the four scenes we shoot

+ APPENDIX A: TOOLS OF THE TRADE →

Today's Bell Schedule

// also the shooting schedule, one period per scene

★ Extra Credit: Tools of the Trade →

Period 1: Homeroom

The Sovereign Stack

A Personal Data Center is the hardware and software you own and run yourself. Instead of renting compute from someone else's cloud, the box sits where you live. You hold the keys, the data, and the off switch.

OP 01

// HOMEROOM: THE SOVEREIGN STACK

Build It From Parts You Own

A Personal Data Center is the hardware and software you own and run yourself. Instead of renting compute from someone else's cloud, the box sits where you live. You hold the keys, the data, and the off switch.

  • The box. A Start9 server is the hardware. A small machine you own outright.
  • The OS. StartOS packages and runs software as services. Install one the way you'd install an app, except it runs on your hardware.
  • Bespoke software. Build the exact app you need, package it as a StartOS service (.s9pk), and install it next to everything else.
  • PWA access. Open the service on your phone or laptop, pin it to the home screen. Native feel, your box behind it.
  • Reach, privately. Keep it local, or publish a single service to the open web through Start-Tunnel, a reverse tunnel through a VPS. The world sees the tunnel's address. Your home IP stays hidden.
  • Security you can verify. StartOS sandboxes every service, encrypts your data, and exposes nothing by default. No shared tenancy, no attack surface you don't control. Self-hosting with Start9 doesn't mean giving up security. It means knowing exactly what's running and who has access.
startos // field readout
# build the service, package it, sideload it
$ github actions → mempoolmadness.s9pk
✓ package built in CI
$ start-cli package install mempoolmadness.s9pk
✓ service running on your box
# expose it through Start-Tunnel, your call
→ tunnel fronts a VPS, visitors see its IP
→ your home address stays hidden
✓ reachable on the open web, on your terms

LEAVE NO TRACE

Start-Tunnel punches a reverse tunnel from your server to a cheap VPS. Incoming traffic hits the VPS first and gets relayed back through the tunnel. The address the world sees is the VPS, not your house. You pick which services go public. Everything else stays behind your firewall.

// reverse tunnel, VPS out front, home server behind it

TRUST BUT VERIFY

The common objection: self-hosting is less secure than leaving it to a cloud provider with a security team. But a Start9 server has a tiny attack surface. No shared tenancy, no exposed admin panels, no third-party access. Every service runs sandboxed. Your data is encrypted at rest. You're not trading security for sovereignty. You're trading blind trust for something you can actually verify.

// smaller footprint, fewer attack vectors, zero shared tenancy

Hall Pass: your home IP stays off the record

Start-Tunnel punches a reverse tunnel from your server to a cheap VPS. Incoming traffic hits the VPS first and gets relayed back. The address the world sees is the VPS, not your house. You pick which services go public. Everything else stays behind your firewall.

// reverse tunnel, VPS out front, home server behind it

Hall Pass: self-hosting doesn't mean giving up security

A Start9 server has a tiny attack surface. No shared tenancy, no exposed admin panels, no third-party access. Every service runs sandboxed. Your data is encrypted at rest. You're not trading security for sovereignty. You're trading blind trust for something you can actually verify.

// smaller footprint, fewer attack vectors, zero shared tenancy
Period 2: A.I. Lab

Spec, Build, Iterate

AI is a force multiplier. One person with the right tools can spec, build, and ship production software. You drive a coding agent for the heavy builds. Your always-on agent runs 24/7 on your server and handles the iteration: specs, fixes, monitoring. Self-hosting that capability means no platform can throttle it or take it away.

OP 02

// FIELD LAB: THE BUILD LOOP

Spec, Build, Iterate

AI is a force multiplier. One person with the right tools can spec, build, and ship production software. You drive a coding agent for the heavy builds. Your always-on agent runs 24/7 on your server and handles the iteration: specs, fixes, monitoring. Self-hosting that capability means no platform can throttle it or take it away.

  • Thought dump → working software. Describe what you want in plain language. A rough idea, not a formal spec. You drive a coding agent (Claude Code, Cursor, Windsurf, pick your tool) and it turns the idea into working software. What used to take a team and a quarter takes one person and a weekend.
  • Sideload and run. CI builds the .s9pk. You install it on your Start9 server. It's running in minutes, on hardware you own, accessible from anywhere.
  • Your always-on agent iterates. An AI agent running on a harness like OpenClaw lives on your server 24/7. It monitors services, manages memory, writes specs for the next coding round, and pushes smaller fixes directly through self-hosted Git. It's the layer between you and the build loop.
  • Three compute layers. Cloud APIs for the best reasoning. Local GPUs for inference, sensitive analysis, and overnight jobs. Your server as the hub that ties them together. No single point of vendor lock-in.
  • Efficiency you own. AI handles the scaffolding, tests, iteration, and monitoring. You make the decisions. Self-hosting that capability on your own hardware means no vendor can change terms, throttle access, or cut you off. The productivity gain is yours to keep.
the build loop
# LAYER 1: you + coding agent (cloud API)
"I want a daily newspaper that
 calls our fine-tuned model and publishes"
youclaude-code → builds, tests, packages
✓ LLaMA-Factory.s9pk ready

# LAYER 2: your server (the hub)
$ start-cli package install -s LLaMA-Factory.s9pk
✓ service running on your box

# LAYER 3: always-on agent iterates
agent → monitors, specs, fixes → gitea
agent → routes work to local GPUs or cloud
agent → daily backups → gitea
cloud APIs for reasoninglocal GPUs for inferenceserver as the hubyou hold the merge
Period 3: Computer Class

Your Own Weights

The model itself can be yours. Fine-tune an open model on your own writing, on your own GPU, and use it to produce real published work. The Daily Vol: a bitcoin and macro newspaper generated every morning by a model trained on your editorial voice. Nothing leaves the building.

OP 03

// THE WORKBENCH: OPEN SOURCE ON YOUR BOX

Your Own Weights

The model itself can be yours. Fine-tune an open model on your own writing, on your own GPU, and use it to produce real published work. The Daily Vol: a bitcoin and macro newspaper generated every morning by a model trained on your editorial voice. Nothing leaves the building.

  • Open source, on your box. No license server, no phone-home. The code runs because you run it.
  • LLaMA-Factory on DGX Spark. A fine-tuning orchestrator packaged for StartOS. It dispatches training jobs to your GPU hardware over SSH. Pick your base model, point it at your data, go.
  • Fine-tune on your own writing. Train an open model on your corpus. In this case, 181 issues of a bitcoin newsletter. It learns your editorial voice. The result: a model that doesn't just answer questions. It writes like you.
  • Publish The Daily Vol. A daily bitcoin and macro newspaper generated entirely by the fine-tuned model, served from your own box. Proof that homebrew weights can produce real published work. No API, no subscription, no per-token cost.
llama-factory // train & publish
# fine-tune on your own writing, your own GPU
$ llamafactory-cli train config.yaml
  model:   google/gemma-4-31B-it
  data:    181 newsletter issues (LoRA, 3 epochs)
✓ training complete, single DGX Spark, 128GB

# serve the fine-tuned model on your LAN
$ vllm serve ~/models/gemma-4-31B-ten31-v2
✓ OpenAI-compatible endpoint, zero API cost

# the result: a daily bitcoin newspaper
→ The Daily Vol, generated, formatted, published
  your model, your hardware, your editorial voice
Period 4: After School

Work While You Sleep

Cron turns the PDC from a tool you poke into a system that runs itself. Your always-on agent schedules jobs against your local models: research, generation, data pipelines. It publishes the results. Share selectively through Start-Tunnel without giving away the box.

OP 04

// AFTER HOURS: SET IT AND WALK AWAY

Rig It to Run Itself

Cron turns the PDC from a tool you poke into a system that runs itself. Your always-on agent schedules jobs against your local models: research, generation, data pipelines. It publishes the results. Share selectively through Start-Tunnel without giving away the box.

  • Cron + your own models. Your always-on agent manages the schedule: newspaper generation, data pipelines, workspace backups, health checks. All hitting models on your LAN. No per-token meter running.
  • A real example: The Daily Vol. Every morning at 3am, a cron job wakes the fine-tuned model, generates a full newspaper edition, formats it, and publishes it to the web. All before you open your eyes. A production workflow on hardware you own.
  • The agent orchestrates. Your always-on agent doesn't just run cron. It monitors the pipeline, swaps vLLM models when needed, falls back to alternate weights if the primary is busy, and logs everything to self-hosted Git.
  • Share selectively. Publish TheDailyVol.AI to the open web through Start-Tunnel while keeping everything else behind your firewall. The GPUs, the agent, the Git server all stay behind it.
crontab // the overnight shift
# 3am CST: your agent fires the daily pipeline
cron → swap vLLM to fine-tuned weights
     → generate newspaper (local GPU, no API)
     → format the edition
     → publish TheDailyVol.AI
     → swap vLLM back to base model
✓ runs every morning, on your hardware

# what's public vs. what stays private
→ TheDailyVol.AI: reachable via Start-Tunnel
→ GPUs, agent, Git: behind your firewall
→ Your home IP: hidden
scheduledself-hosted inferencepublish on your termsno per-token meter
Extra Credit

Tools of the Trade

Every rig comes down to six things across three layers: the server at the hub, GPUs for local compute, and cloud APIs for heavy reasoning. Plus a VPS for the tunnel, the open weights, and the software glue. Here's my kit, and the field substitutes.

APPX A

// REQUISITIONS: WHAT'S IN THE BAG

Tools of the Trade

Every rig comes down to six things across three layers: the server at the hub, GPUs for local compute, and cloud APIs for heavy reasoning. Plus a VPS for the tunnel, the open weights, and the software glue. Here's my kit, and the field substitutes if you're working with what's on hand.

KIT 01 · THE BOX

Hardware StartOS runs on. Your always-on server.

MY KITStart9 Server One. StartOS preflashed, plug in and go. US: Solo Satoshi.

Field substitutes

  • Start9 Server Pure (fully open firmware) or Server Lite
  • Any x86-64 mini PC: Intel NUC, Beelink, Minisforum
  • Refurb business desktop: Dell OptiPlex, Lenovo ThinkCentre
  • A VM via Proxmox. StartOS ships x86-64, ARM64 & RISC-V images

KIT 02 · THE GPUS

The silicon that fine-tunes and serves your own models.

MY KITDual NVIDIA DGX Spark, linked over ConnectX-7 at 200 Gbps, 128GB unified memory each. One serves inference, the other handles training. Both orchestrated from StartOS over SSH.

Field substitutes

KIT 03 · THE VPS

The public node Start-Tunnel fronts, so your home IP stays hidden.

MY KITDigitalOcean. A cheap droplet is all the tunnel needs.

Field substitutes

KIT 04 · THE AI TOOLS

A coding agent for heavy builds, an API agent for specs and iteration.

MY KITClaude Code for heavy builds (I drive it directly) + Anthropic API (via an AI harness like OpenClaw) for always-on orchestration, specs, and iteration.

Field substitutes

KIT 05 · THE MODELS

Open weights you download, host, and fine-tune yourself. Nothing leaves the box.

MY KITGemma 4 (base + fine-tuned) via vLLM, Qwen 2.5 72B via Ollama for the local-only tier. All pulled from Hugging Face and run on the Sparks.

Sources & substitutes

KIT 06 · THE SOFTWARE STACK

The glue that ties hardware to services. An OS to manage it, an agent to automate it, an inference server to run your models.

MY KITStartOS + OpenClaw (AI harness) + Gitea + vLLM + Ollama + LLaMA-Factory

Field substitutes