claudetools/wiki/projects/wrightstown-smarthome.md

---
type: project
name: wrightstown-smarthome
display_name: Wrightstown Smart Home
last_compiled: 2026-05-24
compiled_by: DESKTOP-0O8A1RL/claude-main
sources:
  - projects/wrightstown-smarthome/session-logs/2026-02-09-session.md
---

# Wrightstown Smart Home

## Overview

A privacy-first smart home automation project for the Wrightstown residence. The goal is a fully local-first home automation system with no dependency on Google, Amazon Alexa, or other cloud platforms for core functionality. A hybrid LLM bridge allows selective use of cloud AI for reasoning and search while keeping private data (cameras, sensors, presence) entirely local.

- **Status:** Planning / initial research phase as of 2026-02-09. No hardware deployed yet.
- **Scope:** Home Assistant Yellow setup, local LLM server build, hybrid AI bridge, VLAN network security.
- **Related project:** [[wrightstown-solar]] — planned future crossover via Victron Modbus TCP integration.

## Tech Stack

- **Home automation platform:** Home Assistant (open source, local-first)
- **Hardware:** Home Assistant Yellow (already owned, CM4/CM5 module needed)
- **Connectivity:** Built-in Zigbee 3.0 radio; M.2 NVMe slot
- **Local voice:** Wyoming + Whisper + Piper (fully local, no cloud)
- **LLM inference:** Ollama (primary local runtime)
- **LLM interface:** Open WebUI
- **LLM routing layer:** LiteLLM proxy (unified API, cost tracking, fallbacks)
- **Cloud AI — reasoning:** Claude API (Anthropic)
- **Cloud AI — search:** Grok API (xAI, 2M context, internet access)
- **HA integration:** Extended OpenAI Conversation (connects HA to LiteLLM)
- **Remote access:** Tailscale or WireGuard [unverified — not yet selected]

## Architecture

### Home Automation Core
- Home Assistant Yellow as the central hub
- Zigbee 3.0 radio built-in for device connectivity
- All automations, sensors, and local voice run entirely on-device

### Hybrid LLM Bridge
- LiteLLM proxy as unified routing layer with OpenAI-compatible API
- Estimated routing split: 80% local (Ollama) / 15% Claude API / 5% Grok API
- Estimated cloud cost: ~$5/month
- Routing progression: manual → keyword-based → semantic (eventual)
- PII sanitization pipeline for any cloud-bound queries
- Private data (cameras, sensors, presence) stays local only

### Local LLM Server (Planned Build)
- **Recommended GPU:** RTX 4090 24GB ($1,940–2,240)
- **Alternative builds researched:** budget build (~$580), flagship ($4,000+), Mac Mini M4
- **Primary model (fast/voice):** Qwen 2.5 7B
- **Primary model (reasoning):** Llama 3.1 70B Q4

### Network Security
- 4-VLAN architecture: Trusted / Infrastructure / IoT / Guest
- IoT VLAN isolation: devices cannot reach trusted network
- VLAN hardware decision pending: TP-Link Omada vs Ubiquiti UniFi [unverified]

## Deployment / Hosting

- Runs entirely on-premises at Wrightstown residence
- No cloud hosting; cloud APIs used selectively via LiteLLM proxy
- Remote access via Tailscale or WireGuard (not yet configured)

## Configuration / Credentials

- **Claude API key:** [unverified — account not yet created as of 2026-02-09]
- **Grok API key:** [unverified — account not yet created as of 2026-02-09]
- When created, store in SOPS vault under `clients/wrightstown/` or `projects/wrightstown-smarthome/`

## Active Work / Open Items

All items were pending as of the 2026-02-09 initial research session:

- [ ] Confirm whether CM4 or CM5 compute module is already owned or needs purchasing
- [ ] Set up HA Yellow (basic install, Zigbee, first automations)
- [ ] Research specific Zigbee devices to purchase
- [ ] Finalize LLM server GPU budget (budget 3060 vs sweet-spot 4090)
- [ ] Purchase and build LLM server hardware
- [ ] Decide on VLAN hardware (TP-Link Omada vs Ubiquiti UniFi)
- [ ] Set up Ollama + Open WebUI
- [ ] Create Anthropic API account + Grok API account
- [ ] Configure LiteLLM proxy
- [ ] Integrate HA with LiteLLM via Extended OpenAI Conversation
- [ ] Plan and implement Victron Modbus TCP crossover with [[wrightstown-solar]]

## Key Events / History

### 2026-02-09 — Initial Research and Planning Session
- Session run on machine: ACG-M-L5090
- Defined project scope: privacy-first, no Google/Alexa, HA Yellow as hub
- Researched and selected all major components (HA Yellow, LiteLLM, Ollama, voice stack)
- Designed hybrid LLM bridge architecture with 80/15/5 routing split
- Designed 4-VLAN network security model
- Researched local LLM server hardware options; recommended RTX 4090 24GB build
- Created project documentation structure:
  - `projects/wrightstown-smarthome/PROJECT_INDEX.md`
  - `projects/wrightstown-smarthome/documentation/ha-yellow-setup.md`
  - `projects/wrightstown-smarthome/documentation/llm-server-build.md`
  - `projects/wrightstown-smarthome/documentation/hybrid-bridge.md`
  - `projects/wrightstown-smarthome/documentation/network-security.md`
- No hardware purchased or deployed this session. All work was research and planning.

## Anti-Patterns / Warnings

- [WARNING] **HA Yellow requires a CM4 or CM5 compute module — it does not include one.** Verify ownership before ordering other hardware.
- [WARNING] **JK BMS CAN pinout is reversed** (noted in [[wrightstown-solar]] research, relevant to future crossover). Use USB-UART path with dbus-serialbattery driver, not CAN direct.
- **Do not use Google Home or Amazon Alexa integrations** — privacy-first constraint is a hard project requirement.
- **PII sanitization is mandatory** before any data leaves the local network to cloud APIs.
- **Cloud API credentials must go in the SOPS vault** when created — do not hardcode in HA configuration files.

## Backlinks

- [[wrightstown-solar]] — Related project at same residence; planned Victron Modbus TCP crossover