🌿 Sprinkify — AI-Optimized Irrigation That Waters What Matters

🧠 Inspiration / Problem

Conventional campus and city sprinklers waste massive amounts of water:

🚿 Overspray: sprinklers water sidewalks, buildings, and pedestrians.
💧 Overlap waste: multiple sprinklers redundantly cover the same spots.
🕒 Rigid schedules: timers don’t adapt to lawn shapes or weather.

This leads to water waste, poor efficiency, and negative user experience — a serious issue in cities and campuses trying to conserve water.

💡 What It Does

Sprinkify is an AI-powered irrigation assistant that:

Accepts text descriptions (via Gemini LLM) or images / aerial footage (via ML + OpenCV) to generate GeoJSON lawn layouts with sprinkler positions.
Runs an optimizer (CP-SAT solver + PPO reinforcement learning agent) to generate efficient sprinkler rotation + flow patterns.
Produces simulation heatmaps, MATLAB-powered visualizations, and drives real hardware (servo motors + pumps) to execute the plan.
Offers a React + Tailwind dashboard with chatbot interface, multilingual support, and analytics.
Uses Spring Boot for secure orchestration and Flask for ML/image processing.

👉 The result: maximum coverage, minimum waste, and real-time synchronized sprinkler control.

🏗️ System Architecture

🖥️ Frontend (React + Tailwind + Konva)

Chatbot UI (Gemini agents: Designer, Personal Advisor, Eco-Optimizer)
Image & drone footage uploader
Lawn editor + simulation viewer

☁️ Backend Layer 1 — Spring Boot (Java)

API gateway
Hosts chatbot orchestration
Uses Gemini API for text → GeoJSON parsing & explanations

🧠 Backend Layer 2 — Flask (Python)

Image processing (OpenCV + Shapely):
- HSV green segmentation → lawn polygons
- Contour extraction → Shapely polygons (scaled to real units)
- HoughCircles → sprinkler positions
Optimizer:
- OR-Tools CP-SAT solver (deterministic)
- PPO RL agent (learning-based)
Simulation Renderer:
- Generates coverage heatmaps & MP4 animations

🗄️ Database — MongoDB Atlas

Stores users, lawns, sprinklers, GeoJSON, optimization plans, simulation runs, device telemetry

🧩 Feature Breakdown

🌿 Feature A — Text → Lawn (Gemini LLM, NLP)

User enters natural language (e.g., “5m wide, 7m long lawn with rounded edges…”)
Gemini extracts shape, size, sprinkler placements
Regex + spaCy validation layer ensures units and geometry are consistent
Stored as GeoJSON in MongoDB

🖼️ Feature B — Image/Drone → Lawn (CV)

HSV threshold green → lawn
OpenCV contours → polygons (Shapely)
Threshold blue + HoughCircles → sprinklers
Scale to real units → GeoJSON
Great for both uploaded photos and drone footage (DJI Mini 3 4K) to map irregular lawns automatically.

🧮 Feature C — GeoJSON → Optimized Plan

Solver Path (OR-Tools):
- Rasterize lawn, precompute spray kernels
- Solve scheduling with CP-SAT → minimize overspray + water use
Learning Path (PPO):
- RL agent learns rotation/flow control over simulated lawns
- Adaptable to new shapes without retraining
Hybrid Mode:
- Seed solver with RL output for optimal + feasible schedules.

Output: JSON rotation & flow instructions per sprinkler.

🌐 Feature F — Database & Data Management

All structured data stored in MongoDB: users, GeoJSON, optimization results, telemetry, MATLAB outputs.

🚁 Drone → GeoJSON Feature

We use DJI Mini 3 (4K) drone to capture aerial footage.
Flask runs an OpenCV segmentation + contour pipeline to extract real-world polygons and generate GeoJSON, even for complex or snake-shaped lawns.
👉 This makes Sprinkify scalable for campus-sized mapping.

📊 MATLAB Integrations

8️⃣ Coverage Heatmap Generation

MATLAB simulates water coverage over 2D grid (GeoJSON + sprinklers)
Produces coverage matrices + heatmap PNGs for the dashboard.

9️⃣ Optimizer Benchmarking

Run MATLAB Optimization Toolbox (fmincon, ga)
Compare solutions from OR-Tools, PPO RL, and MATLAB solver.

🔟 Geometric Preprocessing & Shape Analysis

Convex hull, perimeter, area, polygon decomposition
Used for heuristic sprinkler placement and lawn grid preparation.

1️⃣1️⃣ Simulation Playback Generation

MATLAB animates sprinklers over time (angles, flows)
Generates MP4 with VideoWriter → displayed in React dashboard.

🤖 Hardware Build

A mini irrigation controller built on Arduino-class hardware:

Component	Function
Elegoo Uno R3	Microcontroller
16×2 LCD	Status display
28BYJ-48 stepper + ULN2003	Sprinkler rotation
L293D H-bridge + DC pump	Variable flow control
IR tripwire sensor	Stops watering if beam is broken
Analog soil moisture sensor	Monitors soil hydration
Buzzer	Alerts
5 V buck converter	Power regulation

👉 Receives rotation/flow instructions from backend → executes real synchronized patterns.
👉 Displays status on LCD, stops on sensor break, logs moisture — all while rotating like a real sprinkler.

🧠 Gemini Integrations Throughout

Text → GeoJSON parsing
Three advisor modes: Lawn Designer 🧠, Personal Advisor 👤, Eco-Optimizer 🌍
Explanations of optimizer results (e.g., “Sprinkler S1 rotates 40°–120° to minimize overlap…”)
Sustainability narratives (“This schedule saves 12,000L/year ≈ 3 swimming pools”)
Multilingual support for global users

🧪 Tech Stack Summary

Frontend: React + Tailwind + Konva
Auth: Auth0 (MFA, JWT)
LLM: Gemini API (text→GeoJSON, explanations, narratives)
CV: OpenCV + Shapely
Optimizer: OR-Tools CP-SAT + PPO RL
Analytics: MATLAB (heatmaps, optimization, geometry, MP4)
Backend: Spring Boot (gateway), Flask (ML/CV/optimizer)
DB: MongoDB Atlas
Hardware: Arduino-class controller + DJI drone integration

🚧 Challenges

Synchronizing multiple sprinklers with rotation limits and real-time flow.
Handling informal text input (solved via Gemini + validation).
Cleaning drone footage for reliable polygon extraction.
Making solver + RL outputs feasible and explainable.

🏆 Accomplishments

Full pipeline: text/image/drone → GeoJSON → optimized plan → MATLAB analysis → hardware execution.
Real working controller with sensors, rotation, and flow control.
MATLAB integration for heatmaps, optimizer benchmarking, and animations.
Gemini used for both authoring and explanation, making the AI feel like an advisor.

📚 What We Learned

Combining optimization and learning yields superior plans.
MATLAB accelerates numerical analysis dramatically.
Explainability (Gemini) improves user trust and makes demos shine.
A well-designed multi-agent chatbot UX adds massive perceived intelligence.

🚀 What’s Next

Weather-aware & predictive watering (LSTM/XGBoost)
Edge ML on microcontrollers for pedestrian detection
Multi-node coordination across large campuses
AR visualization of sprinkler coverage
Voice mode (Gemini + ElevenLabs)

🛠️ Built With

React · Tailwind · Konva · Spring Boot · Flask · OpenCV · Shapely · Gemini API · OR-Tools · PPO RL · MATLAB · MongoDB · Auth0 · Arduino · DJI Mini 3 Drone