Architecture

Everything behind the chat box.

Seven layers, from bare metal to this public page. Model names and open-source tools below are real; any network address, hostname, or account detail is intentionally generalized.

How a live demo request actually travels

Visitor

Browses this Vercel-hosted site

Next.js API route

Serverless function, server-side only

Tailscale Funnel

One authenticated public doorway

Home server

Local model runtime + knowledge banks

If the home server is offline or the bridge times out, the site falls back to a pre-written example response instead of showing an error.

Hardware

A single high-memory workstation, not a cloud cluster.

Apple Silicon workstation

Ultra-class chip, 24+ CPU cores, 60+ GPU cores

Unified memory

96 GB shared CPU/GPU pool — the key unlock for large local models

Dedicated storage volume

2 TB+ SSD reserved entirely for models, knowledge banks, logs

Always-on networking

Wired Ethernet + static local address for 24/7 availability

Model Serving

The runtime that turns downloaded weights into a live API.

Ollama

OpenAI-compatible API backbone at /v1 — every tool below speaks to this

LM Studio

Optional GUI for trialing new model releases on Apple's MLX engine

llama.cpp / whisper.cpp

Underlying inference engines for text and speech

Models on Disk

Open-weight models chosen to fit a 96 GB unified-memory budget.

gpt-oss-120b (MoE)

~64 GB class — frontier-level reasoning, fully offline

Qwen3.6-35B-A3B (MoE)

Fast daily-driver: coding, long context, vision

Qwen3-VL-30B-A3B

Image, screenshot, and video-frame understanding

Whisper large-v3-turbo

Speech-to-text for meetings and recordings

nomic-embed-text

Embedding model powering retrieval search

Chat & Access Layer

How humans (and other apps) actually reach the models.

Open WebUI

Multi-user chat UI, logins, per-chat model switching, tool calling

Tailscale mesh VPN

Private remote access from anywhere, zero exposed ports

Tailscale Funnel

The one deliberate, authenticated doorway this public site uses

Knowledge & RAG

Three tiers, added only as each one proves insufficient.

Tier 1 — Vector RAG

Domain knowledge banks indexed for retrieval-augmented answers

Tier 2 — LLM Wiki

Obsidian vault of distilled, human-corrected concept pages

Tier 3 — Graph Knowledge

LightRAG for relationship/"what's affected if" questions

Dev & Automation

Turning a chat server into a platform for real project work.

VS Code + Continue/Cline

Local coding assistant against the workhorse model

Python + FastAPI + ChromaDB

Custom retrieval API for programmatic project use

Per-project AI registry

One config per project: model, endpoint, prompt, knowledge bank

This Showcase Site

The public layer you're looking at right now.

Next.js 14 + TypeScript

App Router, React Server Components

Tailwind CSS

Design system, dark mode default

Neon Postgres

Serverless database backing the setup checklist

Vercel

Hosting, edge network, environment-managed deploys

Tailscale Funnel bridge

Lets this site call the real home server's API, safely

A note on safety

The home model server is never opened to the public internet as a whole — only a single narrow endpoint is bridged out, behind a shared secret, for this one demo feature. Everything else (the admin chat UI, the raw model API, project files) stays private to the home network and its mesh VPN.