$30 Radio + Local AI = Internet-Free Smart Home — A Practical Edge AI Case Study

Overview

When the internet goes down, smart homes stop working. Cloud-based voice assistants and IoT automations all become useless. But a developer has built a fully offline smart home using nothing more than a $30 LoRa radio and local LLMs running on a Mac mini.

Living in Ukraine during the war, where power outages and internet disruptions are frequent due to attacks on the power grid, this developer combined Meshtastic LoRa radios with Ollama local models to create a zero-internet smart home control system.

This article analyzes the project’s architecture, tech stack, implementation costs, and the practical possibilities of edge AI.

System Architecture

The overall system is surprisingly simple.

graph TD
    A[T-Echo Portable Radio] -->|LoRa 433MHz<br/>Encrypted| B[T-Echo USB → Mac mini]
    B --> C{Message Routing}
    C -->|SAY: prefix| D[Home Assistant TTS → Speaker]
    C -->|AI: prefix| E[phi4-mini Classifier → gemma3:12b]
    C -->|Status query| F[Home Assistant Sensors]
    C -->|Online?| G[Discord → Cloud AI]
    C -->|Offline?| H[Ollama Local Models]
    I[Outbox Folder] -->|Auto-send| A

The key is dual routing. When internet is available, it leverages cloud AI. When it’s not, it automatically switches to local models. Users don’t need to notice the difference.

Core Tech Stack

LoRa Communication — Meshtastic + Lilygo T-Echo

Meshtastic is an open-source LoRa mesh network firmware. Since each node relays messages, deploying multiple units can create a communication network spanning several kilometers.

• Hardware: Lilygo T-Echo (~$30)
• Frequency: 433MHz LoRa
• Features: Encrypted channels, USB connection, mesh relay
• Limitation: 200-character per message limit (solved by auto-chunking AI responses)

Local LLMs — Ollama

The core of offline AI is a two-stage model architecture.

Model	Role	Size	Purpose
phi4-mini	Intent classifier	~2B	Determines “smart home command or question?“
gemma3:12b	Response generator	12B	Actual answers and reasoning

By classifying intent with a lightweight model first and calling the larger model only when needed, real-time responses are achievable even on a Mac mini M4 with 16GB.

Home Assistant Integration

Home Assistant handles smart home control and TTS (text-to-speech).

• Light control, sensor readings, presence detection
• SAY: prefix converts radio text messages to speech output on home speakers
• Ukrainian language TTS support

SAY: Привіт, я скоро буду вдома
→ Radio waves → Mac mini → HA TTS → Speaker plays voice message

A fully offline voice messaging system requiring zero internet connectivity.

Implementation Cost Analysis

The most compelling aspect of this system is its cost.

Item	Price	Notes
Lilygo T-Echo × 2	~$60	Base station + portable
Mac mini M4 16GB	~$500	$0 if already owned
Home Assistant	Free	Open source
Ollama + Models	Free	Open source
Meshtastic Firmware	Free	Open source
HA Voice PE Speaker	~$50	For TTS output
Total Additional Cost	~$110	If Mac mini already owned

Without any cloud AI service monthly subscriptions, a one-time $110 investment delivers a complete offline AI smart home.

Practical Lessons from Edge AI

1. The Value of Offline-First Design

This project was born from the extreme circumstances of war, but the value of offline-first design is universal.

• Disaster scenarios: Communication independence during earthquakes, typhoons, blackouts
• Privacy: Voice data never leaves your local network
• Latency: Local processing improves response speed
• Cost: Zero monthly subscriptions

2. Strategic Use of Small Models

The architecture separating phi4-mini (2B) as router and gemma3:12b as executor is an exemplary pattern for leveraging LLMs on edge devices.

graph LR
    A[User Message] --> B[phi4-mini<br/>Intent Classification]
    B -->|Smart Home Command| C[Home Assistant API]
    B -->|General Question| D[gemma3:12b<br/>Response Generation]
    B -->|TTS Request| E[HA TTS → Speaker]

3. Mesh Network Scalability

Since Meshtastic is a mesh protocol, adding nodes extends communication range. The creator’s vision of a neighborhood-scale AI network is a realistic scenario.

• Local LLMs on each node
• Mesh relay for multi-kilometer coverage
• Community AI infrastructure without internet

How to Build It Yourself

Here are the minimum requirements to replicate this system.

1. Hardware: 2× Lilygo T-Echo, Mac mini (or any Apple Silicon Mac), HA-compatible speaker
2. Software: Meshtastic firmware, Ollama, Home Assistant
3. Models: ollama pull phi4-mini, ollama pull gemma3:12b
4. Listener daemon: Connect USB radio via Meshtastic CLI, build Python daemon for message routing
5. HA integration: Control Home Assistant via REST API or WebSocket

The entire stack is open source, so you can write the code yourself or delegate to an AI coding tool.

Conclusion

$30 radio + local AI = internet-free smart home. This equation is simple, but it clearly demonstrates the practical future of edge AI.

AI systems that don’t depend on the cloud are no longer theoretical. With a 16GB Mac mini and a $30 radio, this is something you can build today. As local LLM performance continues to improve, the edge AI × IoT combination is poised to become one of the most practical AI application domains.

References

• Reddit - I plugged a $30 radio into my Mac mini and told my AI “connect to this”
• Meshtastic Official Site
• Ollama Official Site
• Home Assistant Official Site
• Lilygo T-Echo