smart-recycle

• 

  output

This is the "heart" of my project documentation.

Here is a structured "Project Journey" for Smart-Recycle AI:

💡 What Inspired me Living in or near an urban hub like Ghaziabad, the challenge of waste management is visible every day. I noticed two major problems:

1. The Confusion of Segregation: Most people want to recycle but don't know if a specific snack wrapper or bottle is "dry" or "wet" waste.

2. The "Throw-away" Culture: We often discard items that could easily be repurposed. I wanted to build a tool that makes the "right" environmental choice the "easiest" choice by putting an expert's knowledge into everyone's pocket.

🧠 What I Learned

1. Building this project was a deep dive into the modern AI ecosystem:

2. Multimodal AI: I learned how Large Language Models (LLMs) like Gemini can "see" and interpret pixels, not just text.

3. The Power of Streamlit: I discovered how to turn a Python script into a fully functional web app in hours rather than days.

4. API Management: I learned the technicalities of rate limits, quotas, and how to handle server-side errors effectively.

Prompt Engineering: I realized that how i ask the AI (the prompt) is just as important as the model itself—specifically for getting local results like "Ghaziabad recycling methods."

🛠️ How You Built It

1. I followed a modular development approach:

2. Interface Design: Used Streamlit to create a simple file uploader and image previewer.

3. API Integration: Connected the google-genai SDK to the backend to send image data to Google’s Flash models.

4. Prompt Logic: Crafted a system prompt that forces the AI to provide a structured response (Material ID → DIY Ideas → Local Tips).

Security: Integrated API key management (and started moving toward environment variables) to keep the application secure.

⚠️ Challenges I Faced

1. No project is built without hurdles! My main challenges were:

2. The "429" Quota Wall: Early on, the app would crash because I was sending too many requests during testing. I solved this by adding Trigger Buttons so the API only runs when specifically requested.

3. Model Versioning: I faced 404 Not Found errors when older model names were retired. This taught me to stay updated with documentation and use stable model versions like gemini-1.5-flash or 2.0-flash.

4. Edge Case Images: Sometimes the AI would struggle with crushed or dirty items. I had to refine the instructions to ensure it gave the "most likely" material identification even if the image wasn't perfect.

Built With

• googleaistudio