# Owl Alpha — AI Infrastructure Architecture

The vision: Z840 as the AI nerve centre. Claude as the persistent orchestrator. All models, all agents, all tools centralised on one server — accessible from any device, always on, always context-aware.

## The Goal

- Claude has persistent memory across every session from any device (phone, laptop, desktop)
- Every new chat bootstraps in <30 seconds with full context via CLAUDE.md + wiki
- Specialised agents handle domains — SEO, Divi, infrastructure, home automation, design, customer service
- Local Mixture of Experts routes tasks to the right model — no unnecessary API costs
- Everything runs offline — client data never leaves the network (GDPR-friendly)
- The wiki grows as the living brain — updated after every session

---

## Hardware Layer

### Z840 — AI Nerve Centre
- Dual Xeon
- 2x RTX 3060 arriving Saturday (24GB VRAM combined)
- Optional 3rd RTX 3060 = 36GB VRAM total
- **36GB VRAM unlocks**: 70B models fully in VRAM, no CPU offload, full inference speed

---

## Model Layer

### Ollama — Local Model Serving
Serves all local LLMs via unified API. Handles VRAM management, model switching, quantisation.

### Model Router
Claude decides, n8n or config executes. Task → optimal model:

| Task Type | Model |
|-----------|-------|
| Code, debugging, scripting | DeepSeek |
| Reasoning, planning, long-context | Mixtral / Llama 3.3 70B |
| Quick, high-volume, lightweight | Phi-4 / Gemma 3 |
| Image generation | Flux / SDXL |
| Semantic search / embeddings | nomic-embed |

### Models to install
- [ ] DeepSeek (code/logic)
- [ ] Mixtral 8x7B or Llama 3.3 70B (reasoning)
- [ ] Phi-4 or Gemma 3 (lightweight)
- [ ] nomic-embed-text (embeddings)
- [ ] Flux or SDXL (image gen)

---

## Orchestration Layer

### Claude — Orchestrator & Router
- Reads CLAUDE.md + wiki on every session boot
- Routes tasks to the right model or agent
- Interprets results, makes decisions
- Writes findings back to wiki before session ends
- Accessible from any device via Tailscale + Open WebUI

---

## Memory Layer

### Wiki + CLAUDE.md — Persistent Memory
- CLAUDE.md = session bootstrap (who, what, where, current state)
- Wiki = long-term brain (infrastructure, decisions, client info, project state)
- Embeddings via nomic-embed = semantic search across all wiki pages
- Lives on Hetzner VPS, accessible to Z840 and all agents

---

## Access Layer

### Tailscale — Secure Mesh
Phone, laptop, desktop, all servers on one private network. No port forwarding.

### Open WebUI
Already installed. Front door to all models and agents from any browser on Tailscale.

---

## Automation Layer

### n8n — Agent Workflows
Already running. Triggers agents on events, schedules, webhooks. Routes tasks between agents.

### MCP Servers
One tool layer shared by all agents:
- OPNsense (firewall)
- CyberPanel (hosting)
- Cloudflare (DNS)
- Google Drive
- Home Assistant
- MainWP (WordPress)
- Hetzner (VPS)

---

## Agent Layer

| Agent | Domain | Primary Models |
|-------|--------|----------------|
| **Hermes** | Infrastructure monitoring | Phi/Gemma (lightweight, always on) |
| **SEO Agent** | Audits, content, rankings | DeepSeek (technical), Mixtral (strategy) |
| **Divi Agent** | Design & site builds | DeepSeek (code), Flux (images) |
| **HA Agent** | Home Assistant automations | Phi/Gemma (speed) |
| **Design Agent** | Brand assets, imagery | Flux / SDXL |
| **CS Agents** | Customer service (hireable) | Client-trained, local, GDPR-friendly |

---

## Build Order

### Phase 1 — Saturday (GPU day)
- [ ] Install 2x RTX 3060 in Z840
- [ ] Install Ollama
- [ ] Pull first models: DeepSeek, Phi-4, nomic-embed
- [ ] Test inference speed
- [ ] Decide on 3rd GPU

### Phase 2 — Model layer
- [ ] Pull Mixtral or Llama 70B
- [ ] Install Flux/SDXL
- [ ] Configure model router config
- [ ] Wire Ollama into Open WebUI

### Phase 3 — Memory & bootstrap
- [ ] Formalise CLAUDE.md structure
- [ ] Set up embedding pipeline for wiki semantic search
- [ ] Test full session bootstrap from phone

### Phase 4 — Agents
- [ ] Migrate Hermes to Z840 (already on kaburuaibox)
- [ ] Build SEO agent
- [ ] Build Divi agent
- [ ] Build HA agent
- [ ] Build Design agent

### Phase 5 — Customer service agents
- [ ] Design white-label CS agent framework
- [ ] First client pilot

---

## Diagram

See owl-alpha-architecture.jsx (interactive React diagram)

```mermaid
graph TD
    Z840[Z840 — AI Nerve Centre<br/>2-3x RTX 3060] --> GPU[GPU Layer<br/>24-36GB VRAM]
    Z840 --> Ollama[Ollama<br/>Model Serving]
    GPU --> Ollama
    GPU --> ImgGen[Flux / SDXL<br/>Image Gen]
    Ollama --> DeepSeek[DeepSeek<br/>Code & Logic]
    Ollama --> Mixtral[Mixtral / Llama<br/>Reasoning]
    Ollama --> Phi[Phi-4 / Gemma<br/>Lightweight]
    Ollama --> Embed[Embeddings<br/>nomic-embed]
    Claude[Claude<br/>Orchestrator] --> Router[Model Router]
    Router --> Ollama
    Router --> ImgGen
    Claude --> Wiki[Wiki + CLAUDE.md<br/>Persistent Memory]
    Wiki --> Embed
    Claude --> Tailscale[Tailscale<br/>Secure Mesh]
    Tailscale --> OpenWebUI[Open WebUI]
    Claude --> Hermes & SEO & Divi & HA & Design & CS
    n8n --> Hermes & SEO & Divi & HA & Design & CS
    MCP --> Hermes & SEO & HA
```

---

## Notes

- The name Hermes running on Hetzner is unintentional mythology — messenger god as monitoring agent
- Owl Alpha = the overarching AI stack name
- All client data stays on-premises — strong GDPR position for selling CS agents
- Z840 is powerful enough to run everything; 3rd GPU is insurance not necessity