An end-to-end smart lighting system that combines real-time IoT sensing with a Generative AI operations agent (LumaAgent) — capable of controlling hardware, querying historical data, and executing dynamic time-based policies through natural language.
Most IoT lighting systems use static rule-based thresholds — if motion is detected AND light is low, turn ON. Simple. Predictable. Limited.
This project replaces and extends that paradigm with an LLM-powered agent that acts as an intelligent operations layer on top of the hardware:
| Capability | Traditional IoT | This System |
|---|---|---|
| Lighting control | Fixed rules only | Natural language + rules + AI reasoning |
| Analytics | Pre-built Grafana charts | Ad-hoc SQL via text-to-query |
| Scheduling | None | "Turn off in 10 minutes" → real timer |
| Context awareness | None | Remembers conversation context |
| Occupancy override | Sensor must detect motion | User can declare intent: "I'm reading for 2 hours" |
- Motion-based lighting — PIR sensor triggers relay
- Ambient light control — LDR prevents lights turning on in bright rooms
- Timer-based auto-OFF — configurable hold time after last motion
- Dual modes — AUTO (sensor-driven) and MANUAL (command-driven)
- Natural language control — "Turn off the light in 10 minutes"
- Text-to-SQL analytics — "How long was the light on yesterday?"
- SQL injection protection — agent can only run
SELECTqueries - ReAct agent loop — Gemini reasons → calls tool → reasons again → answers
- Live sensor context — agent reads real-time MQTT data before answering
- Dynamic policy engine — scheduled commands via
setTimeout()
- Live sensor cards with flash animations on value change
- One-click relay control (ON / AUTO / OFF)
- AI chat interface with typing indicator and markdown rendering
- Scheduled actions panel
| Layer | Technology |
|---|---|
| Edge Device | ESP8266 NodeMCU (Arduino C++) |
| Sensors | DHT11 · PIR · LDR |
| Messaging | MQTT (mqtt package · dev.coppercloud.in) |
| Flow Processing | Node-RED |
| Database | MySQL (home_automation.sensor_data) |
| Visualization | Grafana |
| AI Model | Google Gemini 2.5 Flash (Function Calling) |
| Backend | Node.js · Express · nodemon |
| Frontend | HTML · Vanilla CSS · Vanilla JS |
project-iot/
├── esp8266_code/
│ ├── main.ino ← ESP8266 firmware (credentials redacted)
│ └── secrets.h.example ← Credential template (copy → secrets.h)
├── node-red/
│ └── flows.json ← MQTT → MySQL pipeline
├── database/
│ └── schema.sql ← sensor_data table definition
├── dashboard/
│ └── grafana.txt ← Grafana panel reference
├── images/
│ └── ...
├── ai_agent/ ← AI Agent layer
│ ├── backend/
│ │ ├── .env.example ← API key / DB credential template
│ │ ├── .gitignore ← Protects real .env from Git
│ │ ├── package.json ← Node project dependencies
│ │ ├── config.js ← Single source of truth for settings
│ │ ├── mqttClient.js ← MQTT client + live sensor state
│ │ ├── dbClient.js ← MySQL connection pool + SELECT security
│ │ ├── agent.js ← Gemini ReAct loop + 4 tool functions
│ │ └── app.js ← Express server + static file serving
│ └── frontend/
│ ├── index.html
│ ├── style.css
│ └── script.js
└── README.md
- Node.js 18+ & NPM
- MySQL running locally with
home_automationdatabase - Node-RED running and pushing data to MySQL
- Gemini API key from aistudio.google.com
# In esp8266_code/
cp secrets.h.example secrets.h
# Edit secrets.h with your WiFi credentials
# Flash main.ino to your NodeMCU via Arduino IDEcd ai_agent/backend
cp .env.example .env
# Edit .env with your GEMINI_API_KEY and MySQL password
npm install
npm run devhttp://localhost:8000
That's it. One server. No separate frontend build step.
| User Query | What the Agent Does |
|---|---|
| "What is the current temperature?" | Calls get_current_sensor_status() → returns live DHT11 reading |
| "How long was the light on yesterday?" | Generates SELECT SUM(...) → queries MySQL → calculates duration |
| "Turn off the light in 10 minutes" | Calls schedule_relay_action("OFF", 10) → sets setTimeout() timer |
| "Was anyone home after 10 PM last night?" | Text-to-SQL on PIR column with time filter |
| "Switch to AUTO mode" | Publishes AUTO to home/relay/control via MQTT |
| "What's the average humidity this week?" | Runs SELECT AVG(HUMIDITY) with WHERE TIMESTAMP >= ... |
| Risk | Mitigation |
|---|---|
| API key exposure | Stored in .env — listed in .gitignore, never committed |
| WiFi credentials in firmware | Moved to secrets.h — listed in .gitignore |
| LLM prompt injection → SQL write | dbClient.runReadQuery() rejects all non-SELECT statements before execution |
| LLM issuing invalid relay commands | set_relay_mode() validates against allowed values (ON/OFF/AUTO) |
| CORS attacks from rogue pages | Frontend served by Express at same origin — no external origin allowed |
- Predictive occupancy modelling (ML on historical PIR data)
- Multi-room scalability via MQTT topic namespacing
- Mobile app with push notifications on energy waste alerts
- Voice interface (Web Speech API → chat endpoint)
- LLM-generated weekly energy usage reports via email
Vaibhav (xyron24)
This project demonstrates a full-stack IoT + GenAI integration — from edge sensing on embedded hardware, through real-time MQTT messaging, to a production-grade AI agent with function calling, SQL analytics, and a live web dashboard. Every architectural decision was made to solve a real problem, not to add complexity for its own sake.