// JONBEATZ_COMMAND_CENTER :: ACTIVE_STATION

JonBeatz / Personal AI Command Center

A private command center, running live.

Built to run local models without leaving the desk.

Local models, zero-VRAM stills, vector memory, and Google ops. One station. Five pipelines.

Explore pipelinesWorkstation truth

Local · Remote-ready · Five pipelines online

Private command-center desk — dual monitors, gold accent light, dark workstation

STATION HUD · LIVE DESK

Scroll
  • Next.js
  • React
  • TypeScript
  • Python
  • PowerShell
  • Hugging Face
  • ComfyUI
  • LiteLLM
  • DeepSeek
  • Qdrant
  • Mem0
  • LM Studio
  • Tailwind
  • PyTorch
  • ngrok
  • OAuth
Selected work

The pipelines, and the decisions behind them.

These are the live systems this station spends the most time on. From zero-VRAM stills to local GPU and Google ops, each lane is a different part of how the command center runs.

01 — image.jonbeatz2026

$ npm run image:gen -- "station HUD"
→ public/media · FLUX.1-schnell · 0 GB VRAM
  1. 1.
    Promptnpm run image:gen
  2. 2.
    HF InferenceFLUX.1-schnell
  3. 3.
    Outputpublic/media · 1024²
  4. 4.
    VRAM0 GB local

Cloud for stills, local for heavy edit/video. The GPU stays free for the station that needs it.

ONLINE

Hugging Face Cloud

Zero-VRAM stills that any session can fire without parking the GPU.

Problem

Local diffusion eats the 16GB card. Agents and Comfy can’t share the same VRAM budget on a workday.

What I built

A serverless FLUX.1-schnell lane via Hugging Face Inference. One npm script, 1024 stills in seconds, outputs land in public/media.

Key decision

Cloud for stills, local for heavy edit/video. The GPU stays free for the station that needs it.

Inference APIFLUX.1huggingface_hubPython
Docs →Operator lane

02 — mem0.personal2026

$ npm run mem0:search -- "session ritual"
→ jonbeatz_personal · qdrant_personal · scoped
  1. 1.
    Querymem0:search
  2. 2.
    Embedlocal sentence model
  3. 3.
    Qdrantjonbeatz_personal
  4. 4.
    LLMqwen3-4b @ :1234

Scope is the product. Personal collection only — never write MSC memories from this station.

ONLINE

Mem0 Personal Memory

A private vector store so agents pick up where the last session left off.

Problem

Chat context evaporates. Shared project memory would bleed MSC into a personal profile.

What I built

Isolated Qdrant + local embeddings under jonbeatz_personal. Search/add scripts, LM Studio on :1234 for infer.

Key decision

Scope is the product. Personal collection only — never write MSC memories from this station.

QdrantLM Studioqwen3-4b-instructsentence-transformers
Docs →Memory rail

03 — deepseek.proxy2026

$ curl http://127.0.0.1:4000/health
→ deepseek-v4-pro · think off · Cursor-ready
  1. 1.
    CursorOpenAI-shaped client
  2. 2.
    LiteLLM127.0.0.1:4000
  3. 3.
    DeepSeekv4-pro · think off
  4. 4.
    ngrokoptional remote

Thinking disabled for Agent mode. Cost and reliability beat the old $100+ Vertex path.

ONLINE

DeepSeek AI (LiteLLM)

Paid brain on local rails — Cursor-compatible, Vertex retired.

Problem

Cloud Vertex burn was unsustainable. Agents need a stable OpenAI-shaped endpoint with thinking off.

What I built

LiteLLM proxy on :4000, DeepSeek V4 Pro, optional ngrok for remote Cursor. Status + billing scripts in-profile.

Key decision

Thinking disabled for Agent mode. Cost and reliability beat the old $100+ Vertex path.

LiteLLMDeepSeek V4NGROKCursor Agent
Docs →Paid brain rail

04 — comfy.local2026

$ npm run comfy:start:qwen
→ :8188 · CUDA 12.8 · stop frees VRAM
  1. 1.
    Startcomfy:start
  2. 2.
    LoadGGUF / turbo graph
  3. 3.
    Run:8188 · CUDA 12.8
  4. 4.
    Stopcomfy:stop · free VRAM

Start only when asked; stop when done. Never leave VRAM parked overnight.

ONLINE

ComfyUI Local GPU

The workstation lane for turbo stills, edit, upscale, and short video.

Problem

Cloud stills don’t cover inpaint, face restore, or local video. Need an on-box pipeline with a clean start/stop ritual.

What I built

Shared ComfyUI at H:\AI_Models\ComfyUI on :8188. RTX 5060 Ti 16GB, GGUF turbo, CogVideoX/SVD, YOLO assist.

Key decision

Start only when asked; stop when done. Never leave VRAM parked overnight.

CUDA 12.8GGUFCogVideoXSVD-XTYOLO
Docs →GPU lane

05 — gws.oauth2026

$ npm run google:doctor
→ Gmail · Calendar · Drive · token on disk
  1. 1.
    OAuthdesktop consent
  2. 2.
    ScopesGmail · Calendar · Drive
  3. 3.
    Tokenhermes/google_token
  4. 4.
    Agentplain-English ops

Token stays on disk under hermes. Scripts talk; chat never holds the secret.

ONLINE

Google Workspace API

Inbox, calendar, and Drive in operator English — no secrets in chat.

Problem

Manual Gmail triage wastes sessions. Agents need scoped OAuth, not pasted tokens.

What I built

Desktop OAuth for Gmail, Calendar, Drive. Summaries and drafts via the local LiteLLM path.

Key decision

Token stays on disk under hermes. Scripts talk; chat never holds the secret.

OAuth2LiteLLMngrokGoogle GCP
Docs →Ops lane

Also shipped

  • 06

    Hermes Desktop + Telegram

    Away / overseer channel

    PM lane · 2026

    Away-from-desk channel. Cursor builds; Hermes oversees. Vault bridges keep the desk quiet.

    HermesTelegramVault
  • 07

    Vader Vault

    Cross-project memory

    H:\Vader_Vault · durable memory

    Sessions, decisions, gotchas — link, never duplicate TRUTH.

    ObsidianMCPSessions
  • 08

    Fleet sync

    Hub → siblings

    shared-profile-content

    Rituals, skills, and docs propagate without drift.

    PowerShellfleet:syncdocs
How I build

Start with the problem, not the spec.

Bring the constraint and the goal. I dive into the operator workflow, isolate the core lane, and ship tangible value before stacking chrome on the rest of the station.

  1. 01

    Start with the station, not the feature

    Map the operator workflow first — what must stay online, what can wait, what burns VRAM. Specs come after the constraints.

  2. 02

    Isolate the core lane

    One pipeline that earns its keep. Image, memory, paid LLM, local GPU, or Google — ship that lane before stacking chrome.

  3. 03

    Build with leverage

    Cursor, Hermes, and GrokBot draft and research. Architecture, reviews, and go/no-go stay with the operator.

  4. 04

    Keep the system coherent

    TRUTH, ReCall, vault hubs, and Mem0 scopes stay aligned. New tools join the fleet — they don’t fork identity.

  5. 05

    Ship, learn, iterate

    Static export to jon-beatz.com, local dashboard on :3000. Reality teaches faster than another planning doc.

Station atmosphere stillLOCAL · HERMES HUB

Station & knowledge

Personal AI Command Center

Hermes fleet · 2025 — Present

RTX 5060 Ti · 16GB VRAM

ComfyUI · LM Studio · LiteLLM · Mem0

Knowledge rails

TRUTH.md · START-HERE.md · MASTER-COMMANDS.md · Agent-Runbook.md · ReCall.md

About

I'd rather own the whole station.

This is a personal AI playground and automation hub — not a freelance hire page. I build focused, domain-specific systems where the operator workflow matters as much as the model.

I stay involved from the first constraint through architecture, interface, implementation, and deploy. Cursor builds, Hermes oversees, GrokBot researches, and I operate. The workstation is the product.

The goal is a coherent private stack: local models, zero-VRAM stills, vector memory, and Google ops under one command center — scoped, stoppable, and ready for the next session.

Where the hours went.

  • Operator · Hermes fleet

    2025 — Present

    JonBeatz hub · Windows workstation

    Owns the personal AI command center: session rituals, Mem0, DeepSeek rails, ComfyUI, and fleet sync across sibling profiles.

  • Builder · Cursor + Draven

    Ongoing

    Local · command & control

    Implements sites, agents, and tooling. GrokBot specialists research and brief; Cursor ships the code.

  • Creative systems

    Ongoing

    Image · 3D · motion

    Zero-VRAM stills, local Comfy edit/upscale, scroll/3D experiments, and design systems for client and personal surfaces.

Technical depth

What I reach for, and why.

Not a proficiency chart. Four areas, each with the principles that shape how this station uses the tools.

AI & agents

The model is one part of the station. Context, tool boundaries, VRAM, and failure modes matter as much as the prompt.

  • DeepSeek V4
  • LiteLLM
  • Mem0
  • Qdrant
  • LM Studio
  • MCP
  • Cursor Agent
  • Hermes Desktop
  • RAG
  • Tool calling

Product & frontend

The interface is part of the system. Clear flows, command-center chrome, and motion that earns its place.

  • TypeScript
  • Next.js
  • React
  • Tailwind
  • CSS modules
  • Three.js
  • GSAP (on-deck)

Local compute

Prefer machines you can stop. GPU work is ritualized: start when needed, free VRAM when done.

  • ComfyUI
  • PyTorch CUDA
  • GGUF
  • FLUX / z-image
  • CogVideoX
  • SVD
  • YOLO
  • RTX 5060 Ti 16GB

Ops & infrastructure

Deployments, tunnels, and docs should be boring. Static where it must be; Node only when the server earns it.

  • Hostinger
  • ngrok
  • PowerShell
  • Google OAuth
  • Vader Vault
  • Telegram gateway
  • fleet:sync
Next

Bring the whole problem to the station.

Tell me what you're trying to run, what's in the way, and what constraints you're working with. A paragraph is enough to start. I'll take a look and map it to the right pipeline.

Open pipelines

Station online · five pipelinesLocal · Remote-ready · Operator desk · Not a hire-me funnel