AI-Driven Waste Management System Development

debrajoy kitet

Completed work

AI Engineer

FastAPI

Supabase

TypeScript

Waste Management

The project is live on https://hackathon-2-0-chi.vercel.app/.

Sample screenshots from the AI-Driven waste management and recycling system are attached as:

Prompt Engineering Workflow

This project was built through an iterative prompt-first development approach where each component—from backend to frontend—was designed by giving precise instructions to an LLM.

Below is a detailed, chronological breakdown of real prompts used to generate the entire platform. These prompts demonstrate clarity of intent, architectural foresight, and rapid prototyping.

PHASE 1: Backend Foundation (LangGraph + Gemini + Firecrawl)

Setup Waste Analysis Workflow with LangGraph

Prompt:

“Build a LangGraph workflow powered by Gemini and Firecrawl to do the following in 3 steps:

Search and extract up to 5 specific waste items from content related to a user query.

Analyze each item to provide category, recyclability, health risks, and reuse ideas.

Generate 2–3 wealth creation project ideas based on analyzed items. Return a WasteQueryState model containing all data.”

Add General Question Shortcut

Prompt:

“In the run method, detect short general questions like ‘What is waste?’ or ‘How to recycle plastic?’ using keywords like ‘what’, ‘how’, ‘why’, etc., or word count <= 6. For such cases, return a quick answer using Gemini without running the full LangGraph pipeline.”

Add Pydantic Models for Workflow

Prompt:

“Define three Pydantic models:

WasteItemInfo with fields name, category, recyclability, health_risk, reuse_ideas.

WealthIdea with title, description, required_materials, estimated_value.

WasteQueryState with query, extracted_waste_items, items_info, wealth_ideas, quick_answer.”

Firecrawl + Gemini Integration

Prompt:

“In _extract_items_step, use Firecrawl to search for articles related to the query, scrape them, then prompt Gemini to extract specific waste items from the content. Extract maximum of 5 waste items listed one per line.”

PHASE 2: Token Optimization and Output Control

Limit Waste Extraction to One Item

Prompt:

“To reduce token usage, limit extracted waste items to only one most relevant item based on the query context. Trim content if necessary and prioritize item that directly answers the query.”

Improve Wealth Idea Output Format

Prompt:

“Gemini cannot return List[T] directly. Wrap the List[WealthIdea] in a new class WealthIdeasResponse with field ideas: List[WealthIdea]. Use this wrapper as the output schema when calling with_structured_output.”

Trim Wealth Ideas to Two

Prompt:

“In _generate_ideas_step, return only the top 2 most relevant wealth-to-income project ideas based on the waste item. Make sure the ideas are sorted by relevance and impact.”

Fix Gemini JSON Output

Prompt:

“Ensure Gemini returns structured JSON format matching WealthIdea model: Include title, description, required_materials (list), and optional estimated_value. Wrap in valid JSON list and remove unnecessary explanations.”

PHASE 3: Frontend UI with TypeScript + Supabase

Mobile-First React App with API Integration

Prompt:

“Build a mobile-first React app using TypeScript. Fetch response from a FastAPI backend that returns extracted_waste_items, items_info, and wealth_ideas. Display them in collapsible sections or animated cards. Ensure responsive layout and smooth scroll.”

Supabase Authentication Setup

Prompt:

“Add Supabase email/password authentication with secure session handling. Include registration, login, logout, and real-time auth state management using Supabase client. Store tokens securely and show/hide content based on auth state.”

Add Dancing Smart Bin Loader

Prompt:

“Create a loading screen with a dancing smart bin that bounces, rotates, and sparkles using Framer Motion. Show this animation when the app is analyzing a query or waiting for a response.”

API Error Handling

Prompt:

“Implement graceful error handling on frontend when the API call fails or returns invalid data. Show a fallback error card and retry option. Ensure UI doesn’t crash.”

BONUS: UX & Aesthetic Prompts

Enhance Result Presentation

Prompt:

“Render items_info and wealth_ideas using animated, colorful cards. Color-code recyclability (e.g., green for recyclable, red for non-recyclable). Animate entry with staggered slide-in effect using Framer Motion.”

Add Fallback Data for Dev

Prompt:

“Add mock JSON responses for items_info and wealth_ideas in case backend is unavailable. Toggle between mock and live data using a development flag.”

Local Development Setup

Follow these steps to run both the backend and frontend locally:

1. Clone the Repository

git clone https://github.com/your-username/RecycloHub.git
cd RecycloHub

2. Backend Setup (FastAPI + Gemini + Firecrawl)

Step into the backend folder:

cd server

Create and activate a virtual environment:

# Windows
python -m venv .venv
.venv\Scripts\activate

# macOS/Linux
python3 -m venv .venv
source .venv/bin/activate

Install dependencies:

pip install -r requirements.txt

Create a .env file:

Create a .env file in the server/ directory and add the following:

GOOGLE_API_KEY=your_gemini_api_key
FIRECRAWL_API_KEY=your_firecrawl_api_key

⚠️ Replace the values with your actual API keys.

Run the FastAPI server:

uvicorn app.main:app --reload

The backend will run at http://localhost:8000

3. Frontend Setup (React + Supabase)

Step into the frontend folder:

cd ../client

Install frontend dependencies:

npm install

Configure Supabase and backend API:

Create a .env.local file in the client/ folder:

VITE_SUPABASE_URL=https://your-project-id.supabase.co
VITE_SUPABASE_ANON_KEY=your-anon-key
VITE_BACKEND_API_URL=http://localhost:8000/api/analyze

⚠️ Replace the Supabase URL and key with your actual credentials from the Supabase dashboard.

Run the frontend dev server:

npm run dev

The frontend will run at http://localhost:5173

4. Test the Full App

Go to http://localhost:5173

Enter a waste-related query like: "I have banana peels and old electronics at home"

Wait for the dancing smart bin animation while the AI processes your request

View the extracted items, recyclability, and creative income ideas

🛠️ Troubleshooting Tips

Ensure both client and server folders have their respective .env files.

Confirm Gemini and Firecrawl API keys are correct.

If CORS issues occur, check FastAPI CORS middleware settings in main.py.

Use mock data toggle on frontend during offline development.

Security and Fault Tolerance

RecycloHub prioritizes data protection, responsible AI usage, and resilience against unexpected failures through the following practices:

Security Measures

Supabase Authentication Secure and modern user authentication using Supabase with support for email/password login and protected user sessions.

Environment Variables Sensitive API keys (e.g., Gemini, Firecrawl, Supabase) are stored securely in .env files and never exposed to the frontend or committed to version control.

Backend API Protection FastAPI backend is structured to validate and sanitize all incoming queries, reducing injection and malformed input risks.

Structured Output with Validation AI-generated outputs are parsed against strict Pydantic schemas (WasteItemInfo, WealthIdea, WasteQueryState) to prevent malformed or malicious data from being returned to users.

Fault Tolerance Strategies

Graceful Error Handling Try-except blocks around all LLM and Firecrawl calls ensure the app doesn't crash on failure. Instead, it falls back with default messages or empty states.

Partial Response Recovery Even if some steps fail (e.g., analyzing one waste item), others continue. The pipeline returns all successful data and logs failures for debugging.

Development Fallbacks Frontend includes mock data toggles to allow testing without a live backend or API keys, ensuring smooth prototyping.

Retry and Logging Support LangGraph's built-in arun_with_retry logic handles retryable tasks for LLMs to reduce timeout or transient errors.

AI Workflow Architecture

RecycloHub leverages a structured, AI-first architecture to intelligently handle user queries using LangGraph, Gemini, and Firecrawl. The workflow is designed to be efficient, fault-tolerant, and cost-effective.

Workflow Flowchart

User Query
   └──> Is it a general question?
            ├── Yes → Answer directly with Gemini → Done
            └── No → Begin structured pipeline:
                    ├── Step 1: Firecrawl retrieves and scrapes relevant articles
                    ├── Step 2: Gemini extracts specific waste items
                    ├── Step 3: Gemini analyzes each waste item (category, risks, recyclability)
                    └── Step 4: Gemini generates 2 targeted waste-to-wealth ideas

Core Technologies

LangGraph: Manages the multi-step stateful workflow with async support

Google Gemini (via LangChain): Handles all structured language tasks

Firecrawl: Fetches and scrapes real-time web content for contextual AI reasoning

Pydantic Models: Enforce schema validation on all AI outputs

Async/Await: Optimized for performance and concurrency across all I/O operations

The workflow ensures each component is isolated, testable, and fault-resilient—making the system highly maintainable and production-ready.

Test Queries & Expected Output

Use the sample queries below to explore RecycloHub's full capabilities. These examples demonstrate both the general Q&A flow and the full AI analysis pipeline.

General Q&A (Quick Answer)

Ask concise questions that require a definition, explanation, or procedure:

What is e-waste?

How do I recycle used batteries?

Why is waste segregation important?

Who is responsible for waste management in Kenya?

Output: A direct, expert response from Gemini with 1–2 paragraph explanation.

Full AI Analysis & Wealth Ideas

Ask contextual or descriptive questions that include a waste type or scenario:

There are many banana peels in my kitchen. What should I do with them?

I have a pile of scrap metal at home. Can I turn it into something useful?

There’s used cooking oil in my restaurant. How can I reuse or sell it?

My compound is full of tyre wastes. Are they recyclable?

Output Includes:

Extracted waste item(s) from content

Detailed analysis per item:

Type (organic, electronic, plastic)

Recyclability and health risks

Practical reuse ideas

Top 2 relevant waste-to-wealth ideas with materials and earning potential

Like this project

Completed work

Posted Aug 4, 2025

Built AI-driven waste management system using LangGraph, Gemini, and Firecrawl.

Likes

Views