• SYSTEM ONLINE CASESTUDY.R&D BARAD-DÛR PERSONAL NOT CLIENT WORK
C:\PJGREEN> open barad-dur.case

Barad-dûr AI platform

A personal homelab R&D stack: hardware I spec'd and built, a 100+ TB Unraid array, and a local operator console for chat, tools, evals, and multi-GPU inference — with approval gates so agents cannot run wild on my network.

This is not a client deliverable. It is the sandbox where I pressure-test agent patterns, local inference routing, and operator safety before I adapt the ideas into paid work.

Personal, non-commercial 100+ TB self-built array Local LLMs + approval-gated tools Evals, benchmarks, watchdog runs

Why it exists

A place to learn faster than cloud rent allows

I wanted full control over storage, GPUs, and services for AI and automation experiments — without shipping homelab internals to a client environment. Barad-dûr is the hardware plus the operator console that sits on top of it.

Honesty boundary: this write-up stays pattern-level. No internal hostnames, credentials, or client-adjacent data. The point is to show how I think about infra, agent safety, and local AI — not to expose the homelab map.

What I built

Hardware underneath, operator console on top

Homelab hardware & storage

Self-built Unraid server with a 100+ TB storage array, Docker service stack, Home Assistant integration, and multi-GPU nodes (GTX + RTX classes) used for local inference routing.

Barad-dûr operator console

FastAPI backend and Next.js UI for local Ollama chat, admin diagnostics, a code workbench, eval/benchmark harnesses, and SQLite-backed threads, settings, and watchdog runs.

Approval-gated agent tools

Homelab health checks, workspace edits, Home Assistant actions, web search, memory, operator KB search, weather/time lookups, GPU status, and host probes — staged until I approve.

Safety & routing

Multi-GPU routing between lighter and heavier nodes; code edits stay staged until an explicit apply step; bounded modes for experimentation without turning the whole LAN into an open tool surface.

What it proves for clients

Infra depth you can trust when I design your systems layer

Clients rarely need a homelab. They do need someone who understands how data, GPUs, agents, and humans share responsibility — and who has actually operated all three together.

Self-hosted fluency

  • Unraid + Docker operations
  • Multi-TB storage planning
  • LAN-safe service exposure

Agent governance

  • Human-in-the-loop tool gates
  • Evals & benchmark loops
  • Staged code / diff apply

Transferable patterns

  • RAG / context layers
  • Routing & degradation
  • Operator docs & runbooks