Connected home with Home Assistant, Zigbee and AI cameras

The challenge

A house that already had devices from three brands and a server on a Raspberry Pi needed a single system that kept working without internet, exposed neither cameras nor dashboards to the internet and gave the resident useful alerts instead of endless recordings.

The objective

Migrate to a more capable local server without losing history, unify the devices on a Zigbee network, deploy an alarm and cameras with local AI, close all external access behind a VPN and document every decision.

The solution

Home Assistant OS on an Intel mini PC with a UPS, migrated from the Raspberry Pi with a step-by-step plan; Zigbee2MQTT with a local dongle and Mosquitto as the MQTT bus; Alarmo with modes, entry and exit delays, a physical remote and critical phone alerts with an action; Frigate with people detection on the integrated GPU (OpenVINO, about 8 ms per frame) and face recognition; short alert descriptions by a generative-AI model, with the cost cut from hundreds of dollars a month to cents; remote access only through Tailscale with HTTPS in the tunnel, two-step login for administrators, a local user for the wall panel and cameras blocked from reaching the internet; dashboards generated by script through the REST and WebSocket APIs; local backups and two external destinations; 27 decision records and a changelog.

The impact

• 32 Zigbee devices from three brands on a single local network, with no vendor cloud
• 24 automations in operation: lighting by presence and time, monitoring and alarm, each recorded with its reason
• Video retention policy defined after measuring 287 GB in two days of continuous recording
• Cost of the AI that describes alerts cut from hundreds of dollars a month to cents
• 27 documented architecture decisions, from hardware to retention rules