A frontier-class AI server
running entirely on one machine.
This site showcases the design, setup, and roadmap for a fully local LLM server — the hardware, the model stack, an interactive build checklist, and where it's headed next: real agentic AI, with guardrails from day one.
Why local
Frontier capability, kept close to home.
One machine, not a cloud bill
A single high-memory workstation runs the entire model stack — no per-token API charges, no data leaving the building.
A real model stack, not one model
A large reasoning model, a fast daily driver, and a vision + speech pair — routed to the right job automatically.
Guardrails before autonomy
Read-only defaults, human approval gates, and full action logging — required before any agent touches real data.
Built to expand
Every future agent — audit, budget, education, content — reuses the same foundation: models, retrieval, tools, evaluation.
The stack
Seven layers, one API.
Hardware
A single high-memory workstation, not a cloud cluster.
- Apple Silicon workstation
- Unified memory
Model Serving
The runtime that turns downloaded weights into a live API.
- Ollama
- LM Studio
Models on Disk
Open-weight models chosen to fit a 96 GB unified-memory budget.
- gpt-oss-120b (MoE)
- Qwen3.6-35B-A3B (MoE)
Chat & Access Layer
How humans (and other apps) actually reach the models.
- Open WebUI
- Tailscale mesh VPN
Track progress
An interactive checklist for the whole build.
Every phase of standing this server up — foundation, model stack, knowledge banks, multi-user access, agentic layer — is tracked as a live checklist, not a static document.
Open the setup guideWhat's next
From one server to a family of agents.
DoD FM Knowledge Agent
A grounded Q&A agent over financial-management regulations, guidance, and internal SOPs — answers cite the source document instead of guessing.
DoD FM Audit Agent
Walks a sample of transactions through a documented test procedure, flags exceptions, and drafts the supporting workpaper narrative for human review.
DoD FM Budget Agent
Turns budget-to-actual variance analysis from a multi-hour spreadsheet exercise into a first-draft narrative a human tightens before it ships.
K-12 Education Agent
A standards-aligned tutoring and lesson-support assistant grounded in curricula and assessment banks rather than open-ended web knowledge.