home/work/warden
Founder · BuildingPrivate

Warden

An offline-first smart-home hub that runs a local large-language model on hardware you own — full natural-language control of your home with no cloud, no telemetry, and nothing about your household leaving the building.

Warden prototype motherboard — top-down view
01

Overview

summary

Warden is an offline-first smart-home hub — a single device that runs a large-language model on hardware you own, and gives you full natural-language control of your home with no cloud dependency, no telemetry, and nothing about your household leaving the building.

It is built on two mature, open technologies: Home Assistant for device orchestration and a locally hosted model served through Ollama for understanding and reasoning. The model, the automations, and the data all live on the hub itself. The design goal is uncompromising — a home that is genuinely intelligent without being a listening post.

Warden — exploded view of the prototype hardware
02

The Problem — The Cloud Is Listening

why

The mainstream smart-home and voice platforms are built on a cloud model, and the industry is moving away from local control, not toward it. In March 2025, Amazon removed the one setting that let certain Echo owners keep voice processing on the device, so that essentially every Alexa request is now sent to Amazon's cloud. Amazon has also acknowledged that it collects information about other devices on a customer's home network — even ones Alexa never touches.

“We use your data to improve the AI” has become the standard justification. In practice, your words become permanent training material and behavioral-profile data. Meta now trains its AI on the public posts and comments of adult users across Facebook and Instagram — and by its own account, people's interactions with Meta AI, meaning their questions and queries, are used to train and improve its models too. OpenAI's own documentation states that, on consumer plans, ChatGPT is trained on users' conversations by default unless they hunt down the opt-out.

And the “we would never misuse it” assurances have a track record. In 2019 the Federal Trade Commission hit Facebook with a $5 billion penalty — the largest privacy fine it had ever imposed — after the Cambridge Analytica scandal exposed how tens of millions of users' personal data had been funneled to a third party. The principle Warden is built on follows directly: the only data that can't be leaked, sold, subpoenaed, or quietly repurposed is the data that never leaves your house.

03

The Market

adoption

This isn't a niche worry — it's a mainstream platform exposed to mainstream surveillance. The U.S. smart-home base is large and compounding, which means the privacy gap Warden closes is widening every year.

103M
U.S. smart homes projected by 2028 — after nine straight years of growth
~75%
Projected U.S. household penetration by 2028
700M
Smart devices in use across U.S. homes by 2025

U.S. smart homes are multiplying

Households running smart-home tech, in millions
30M 60M 90M +63% 63M103M 20232028 (proj.)
Source: Statista Market Insights — U.S. smart-home users, ~63M in 2023 rising to a projected 103M by 2028 (+63%).

Already inside U.S. homes

Share of U.S. households owning each device type, 2022
20% 40% 60% 57%50%41%36% Smart TVVoice asst.Smart appl.Smart light
Source: U.S. smart-home market data, 2022 — most-owned device categories.
04

How Warden Is Different

architecture

Warden is local-first from the ground up — not a cloud product with a local mode bolted on:

  • On-device intelligence — the LLM runs locally via Ollama and Home Assistant orchestrates every device on-prem. Prompts and routines never traverse a vendor's servers.
  • Zero default egress — no telemetry, no analytics beacons, no account required to use the hardware you bought.
  • Owner-controlled — you can inspect, export, and erase everything. Remote access, if you want it, is a self-hosted encrypted tunnel — never a vendor relay.
  • No lock-in — open standards (Matter, Zigbee, Z-Wave, Thread) mean the home keeps working even if a manufacturer disappears or sunsets a cloud.
  • No monetization of you — no subscriptions, no ads, no data sales. The product is the hub, not the household.
05

Architecture

data flow

Every block below sits inside the home. Inputs feed a single hub that holds the model, the automations, and the data; outputs are answered locally or relayed to a caretaker through a tunnel you control. The one path that doesn't exist is the route out to a vendor's servers.

Local-first data flow

Where Warden's data goes — and where it can't
VENDOR CLOUD prompts · audio · behavioral profiles NO EGRESS YOUR HOME · LOCAL-ONLY SENSORS mmWave · motion · contact LOCAL MICS on-device wake word DEVICES Zigbee · Matter · Z-Wave WARDEN HUB Home Assistant Ollama · Local LLM STT · TTS pipeline Encrypted store DASHBOARDS & VOICE answered in-home CARETAKER ALERTS well-being signals WIREGUARD TUNNEL owner-controlled
Diagram is illustrative of the prototype architecture.
06

Prototype Hardware

reference build
Warden — assembled prototype hub, front three-quarter view

A proposed reference build. Warden is hardware-agnostic, but these are the components that make a capable, fully local hub — sized so a quantized model and Home Assistant run comfortably side by side.

Compute

Minimum
Intel N100/N305 mini-PC or a Raspberry Pi 5 (8 GB) — runs Home Assistant plus a small quantized model on CPU.
Recommended
Small-form-factor PC with Ryzen 7 / Core i7 + NVIDIA RTX 4060 (8–16 GB VRAM), or an Apple Silicon Mac mini for strong performance-per-watt.

Memory

Minimum
16 GB RAM.
Recommended
32 GB RAM — or 24 GB+ unified memory on Apple Silicon.

Storage

Minimum
512 GB SSD.
Recommended
1 TB NVMe SSD — OS, Home Assistant, models, and logs.

Local model

Minimum
3B–7B, 4-bit quantized (Llama 3.2 3B / Mistral 7B).
Recommended
8B–14B, 4-bit (Llama 3.1 8B / Qwen2.5 14B), served by Ollama.

Smart-home radio

Minimum
Zigbee / Thread / Matter stick (Home Assistant Connect ZBT-1).
Recommended
Add a Z-Wave controller (Aeotec Z-Stick) for full device coverage.

Voice (local)

Minimum
One USB far-field mic array (ReSpeaker).
Recommended
ESP32-S3 voice satellites (Wyoming) for whole-home wake word; openWakeWord + faster-whisper + Piper.

Camera-free sensing

Minimum
PIR motion + door / window contact sensors.
Recommended
Add mmWave presence sensors (Aqara FP2 class) and temp / humidity / air-quality.

Power resilience

Minimum
600 VA UPS.
Recommended
1000–1500 VA UPS with USB monitoring (NUT) for clean ride-through of outages.

Networking

Minimum
Wi-Fi 6.
Recommended
Gigabit Ethernet to the hub; Wi-Fi 6 for satellites.

Software stack — open-source and self-hosted end to end:

  • Home Assistant OS (or HA Supervised on Debian) as the orchestration layer.
  • Ollama serving a 4-bit quantized local LLM.
  • Wyoming voice pipeline — openWakeWord, faster-whisper (speech-to-text), and Piper (text-to-speech), all on-device.
  • Zigbee2MQTT / ZHA plus a Matter server over a local MQTT broker.
  • WireGuard for optional, owner-controlled remote access — never a vendor relay.
07

Features

capabilities

Local voice & text control

An on-device wake word and a locally hosted LLM handle natural-language commands and conversations. No audio, no transcript, and no prompt ever leaves the hub.

No-camera ambient sensing

Presence, motion, door/contact, and environmental sensors instead of video feeds — situational awareness without putting a lens inside someone's home.

Well-being summaries & alerts

Plain-language daily check-ins and anomaly alerts ("routine looks normal today," or a heads-up when it doesn't) delivered to a trusted caretaker.

Offline-first automations

Lighting, climate, locks, and energy routines that keep running when the internet — or the vendor — is gone. Offline is the design target, not a degraded fallback.

Encrypted local storage

Household data lives on the device, encrypted at rest. You can see exactly what exists, export it, and wipe it. There is no copy on someone else's server.

Owner-controlled remote access

Optional self-hosted encrypted tunnel for checking in from afar — your relay, your keys, no third party sitting in the middle of your home.

Resilience hub

UPS / battery-aware and built to keep core safety and comfort functions alive through power and connectivity outages.

Open-standard devices

Matter, Zigbee, Z-Wave, and Thread support — so the system isn't tied to one manufacturer's cloud or bricked when a product line is sunset.

08

Who It's For

applications

Privacy-first households

People who refuse the premise that a smart home has to be a listening post — and who want their prompts, routines, and daily life kept strictly confidential.

Elder care & aging in place

Dignified, camera-free monitoring that reassures family a parent is moving through their day normally, without turning their home into a recorded space.

Families & child safety

Keep a child's routines, footage-free presence data, and the home's "memory" local — not sitting on a corporate server tied to an advertising business.

Outage & off-grid resilience

Cabins, farms, rural and remote properties where cloud assistants simply stop working when the connection drops. Warden also acts as a backup automation brain during power or internet loss.

Confidential professions

Lawyers, clinicians, therapists, and accountants who can't feed privileged or regulated client information into a cloud AI — but still want a capable assistant.

Home offices & small business

Handle customer data, trade secrets, and IP with an assistant engineered so none of it is ever shipped to a vendor for "improvement."

High-exposure individuals

Executives, public figures, journalists, and security-conscious users for whom an always-on cloud microphone is an unacceptable attack surface.

Accessibility that can't fail

Voice control for users with mobility or vision needs that has to keep working even when the internet doesn't — reliability as an accessibility requirement.

09

Sources

receipts

The privacy case for Warden isn't hypothetical. It's drawn from the public record of how the dominant platforms actually treat household data:

»

More Work

projects