Skip to content
James Foss, jamesfoss.me home
P.04Private system· Private infrastructure

Self-Hosted AI Infrastructure

A locally hosted AI environment incorporating NVIDIA GPUs, Ollama, large language models, and AI workflows.

Category
Local LLM platform
Status
Private system
Core stack
Ollama · Local Large Language Models · GPU Computing · Model Deployment · AI Agent Workflows
Fig. 04 — Illustrative cover artwork, not a product screenshot
01

Overview

This is the private AI environment behind Linear Web Solutions. Models run on local NVIDIA GPUs through Ollama, so prompts and data stay on hardware I control.

Two interfaces sit on top of the model layer: a customized fork of Open WebUI, branded as LWS AI, for conversation, and a Langflow deployment, branded as LWS Agents, for designing agent workflows visually.

  • 01Ollama serving models from a dedicated GPU host
  • 02A customized Open WebUI fork as the conversational interface
  • 03Langflow for visual agent and workflow design
  • 04Upgrade-safe customization that never patches installed packages
02

Objective

Run capable language models privately on local hardware, and make them usable through polished interfaces and workflow tools as a private alternative to hosted AI platforms.

03

Technical approach

  1. A.

    Separated model serving

    Ollama runs on its own GPU host and the interfaces connect to it over the network, so the model layer can be upgraded independently of everything that uses it.

  2. B.

    Maintainable forks

    Every change to Open WebUI lives in dedicated files or small, marked hunks, so upstream releases merge cleanly and license and attribution notices stay intact.

  3. C.

    Non-invasive customization

    Langflow's branding is applied to a copy of its compiled frontend, with scripts that apply, inspect, and restore the stock install after every upgrade.

  4. D.

    Isolated services

    Open WebUI ships as a Docker image built on the official release, and Langflow runs in its own Linux container on Proxmox.

04

Architecture

  1. Interfaces

    • LWS AI (Open WebUI fork)
    • LWS Agents (Langflow)

  2. Orchestration

    • Agent and workflow design
    • Conversation and model selection

  3. Model serving

    • Ollama
    • Local language models

  4. Compute

    • NVIDIA GPUs
    • CUDA acceleration

  5. Platform

    • Docker
    • Proxmox LXC
    • Linux

Hostnames, addresses, and access details are intentionally omitted.

05

Technologies

  • Ollama
  • Local Large Language Models
  • GPU Computing
  • Model Deployment
  • AI Agent Workflows
  • Langflow
  • Open WebUI
  • Docker
  • Proxmox
  • Virtualization
  • Networking
  • Linux
  • Python
  • Bash
06

Visuals

Private system. Interface screenshots are not published.

The artwork above is an illustration of the system, not an interface capture.

07

Challenges & solutions

  1. Challenge 01

    Customizing open-source tools without turning every upgrade into a merge project.

    Solution

    Changes are isolated by design: marked hunks and dedicated directories in the Open WebUI fork, and a re-runnable branding kit for Langflow that leaves the installed package untouched.

Have a system like this in mind?

Get in touch