Cloud computing: Send data to AWS, wait for response.
Edge computing: Process data on-device, instant response.

For our SmartHome Hub, edge computing was the only option.

Why Edge Matters

Imagine this: You say "Alexa, lights off." But your internet is down.

Cloud-only system: Lights stay on. (Command requires AWS server)
Edge system: Lights turn off. (Processed locally)

The Architecture

Hardware

• Raspberry Pi 4 (8GB RAM, quad-core)
• Zigbee coordinator (USB)
• Z-Wave stick (USB)
• 1TB SSD (video storage, faster than SD card)

Software Stack

• Home Assistant (open-source hub)
• MQTT broker (Mosquitto)
• Node-RED (automation engine)
• PostgreSQL (local database)

What Runs on the Edge

1. Device Control

Turn lights on/off, adjust thermostat, lock doors—all without internet.

MQTT message flow:

[Light Switch Press] 
  → MQTT: zigbee/button/click
  → Node-RED: IF button=1, THEN lights/living_room/set ON
  → MQTT: zigbee/light/living_room/set {"state":"ON"}
  → [Lights Turn On]

Latency: 32ms (vs 200-500ms cloud round-trip)

2. Video Processing

Security cameras stream to the Pi, not the cloud.

Motion detection:

import cv2

cap = cv2.VideoCapture('rtsp://camera.local/stream')
while True:
    ret, frame = cap.read()
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    
    if motion_detected(gray):
        save_clip(frame)
        send_alert()

Processing: 15 FPS on Pi 4 (good enough for motion detection)

3. Voice Control

Offline speech recognition with Piper TTS + Whisper.

Command: "Turn off bedroom lights"
Processing:

1. Microphone captures audio
2. Whisper transcribes to text (runs on Pi)
3. Intent parser extracts action ("turn off") + entity ("bedroom lights")
4. MQTT command sent locally

No cloud, no subscription.

Cost Comparison

Cloud-based smart home (Nest, Ring):

• Hardware: $500
• Subscriptions: $10-30/month
• 5-year cost: $1,100-2,300

Edge-based (our setup):

• Raspberry Pi 4: $75
• Zigbee/Z-Wave sticks: $60
• SSD: $100
• Smart devices: $300
• Total: $535 (one-time)

Savings over 5 years: $565-1,765

Performance Metrics

Automation response time:

• Motion sensor → Light on: 28ms
• Button press → Scene activation: 41ms
• Voice command → Action: 180ms

Compare to cloud:

• SmartThings: 200-800ms
• Google Home: 300-1,200ms

The MQTT Architecture

Topics Structure

zigbee/
  ├─ light/living_room/state
  ├─ light/living_room/set
  ├─ sensor/motion/bedroom
  └─ button/hallway/click

zwave/
  ├─ lock/front_door/state
  └─ thermostat/living_room/temperature

Node-RED Flow

// When motion detected, turn on lights
msg.payload = {
  state: "ON",
  brightness: 50,
  transition: 1  // 1 second fade
};
return msg;

Security on the Edge

1. No Cloud = No Cloud Leaks

Your video feeds never leave your home network.

2. Local Authentication

Access control managed locally (no AWS IAM to configure).

3. VPN Access

Remote access via WireGuard VPN (encrypted tunnel to your home).

4. Firewall Rules

# Block all incoming except VPN
iptables -A INPUT -p tcp --dport 51820 -j ACCEPT  # WireGuard
iptables -A INPUT -j DROP

Scaling Challenges

1. Limited CPU

Pi 4 can handle ~50 devices before slowdowns.

Solution: Multiple Pis (one per floor/zone)

2. Storage Limits

1TB fills up fast with video recordings.

Solution: Auto-delete after 30 days, upload important clips to S3

3. No Redundancy

Pi crashes = system down.

Solution: Failover to cloud (Home Assistant Cloud as backup)

Real-World Deployment

Smart Home (1,800 sq ft):

• 42 devices (lights, sensors, locks, cameras)
• 1,200 automation rules
• Raspberry Pi 4 CPU usage: 18%
• RAM usage: 2.1 GB / 8 GB
• Uptime: 99.8% (3 crashes in 18 months)

When Edge Isn't Enough

Cloud is better for:

• Machine learning inference (too slow on Pi)
• Multi-site deployments (central management)
• Massive data storage (S3 cheaper than SSDs)

Hybrid approach:

• Process locally (lighting, sensors)
• Cloud for heavy lifting (AI, analytics)

The Future: AI at the Edge

We're testing:

• Google Coral TPU (AI accelerator, $60)
• Object detection (person vs pet vs car)
• Face recognition (unlock door for family)

Performance: 30 FPS object detection on Coral vs 2 FPS on Pi CPU

Lessons Learned

1. SD cards fail. Use SSDs.
2. Power matters. Get a quality power supply (official Pi adapter).
3. Cooling helps. Pi 4 gets hot; add a fan or heatsink.
4. Backups are essential. Automate daily backups to NAS.

Want an edge-first smart home?
Build Your SmartHome Hub