EcoNova Guardian: Agentic Waste Sorting Assistant

An interactive Amazon Nova–powered agent that guides users to the correct bin, asks clarifying questions when uncertain, and tracks common sorting mistakes to help communities improve recycling.

Comment

EcoNova Guardian: Agentic Waste Sorting Assistant

Inspiration

People want to sort waste correctly, but most decisions happen in seconds and local rules are often unclear. That leads to contamination in recycling/compost streams and more landfill waste. We built EcoNova Guardian to make sorting easier at the exact moment people need help.

Our goal was to combine real-time camera guidance with practical AI reasoning so users can learn while they sort, not after the fact.

What it does

EcoNova Guardian is a real-time waste sorting assistant powered by Amazon Nova on AWS Bedrock.

It helps users by:

  • Capturing live camera input
  • Classifying items into Waste, Recycling, or Compost
  • Showing quick reasoning and disposal guidance
  • Asking clarifying prompts when uncertain
  • Collecting feedback to improve future decisions

In short: it turns waste sorting from guesswork into an interactive, guided action.

How we built it

Stack:

  • Frontend: Vanilla JavaScript + HTML/CSS
  • Backend: FastAPI (Python)
  • Model Layer: Amazon Nova foundation model via AWS Bedrock
  • Data: Lightweight local storage for feedback and analytics

Architecture:

  1. Capture + Analyze: Camera frames are sampled and sent to the backend only when meaningful visual change is detected.
  2. Nova Inference: Amazon Nova classifies the visible item and returns structured reasoning.
  3. Guided UX: Users see the bin recommendation and can provide correction feedback.
  4. Learning Loop: Corrections are logged and used for analytics and future tuning.

This design let us balance responsiveness, cost control, and usability.

Challenges we ran into

  1. Avoiding false positives from blank or noisy scenes.
  2. Keeping response speed high while reducing model calls.
  3. Making the UX clear when confidence is uncertain.
  4. Handling sequential items smoothly in one camera session.
  5. Keeping local developer setup stable and fast.

Accomplishments that we're proud of

  • End-to-end prototype running with live camera classification
  • Practical integration of Amazon Nova via Bedrock in a real app flow
  • Noticeable reduction in unnecessary model requests through smart gating
  • Clear correction workflow for user trust and continuous improvement
  • A demo-ready, lightweight system that is easy to run locally

What we learned

  1. Real-time AI products succeed when UX and model logic are designed together.
  2. Prompt design and guardrails matter as much as model choice.
  3. A feedback loop is essential for trust and long-term improvement.
  4. Cost-awareness must be designed from day one, not added later.

What's next for EcoNova Guardian: Agentic Waste Sorting Assistant

  1. Expand to mobile-first deployment.
  2. Add multilingual support for broader accessibility.
  3. Deepen analytics for community and municipal insights.
  4. Improve personalization with smarter feedback-driven guidance.
  5. Explore integrations with smart-bin and sustainability programs.

Built with Amazon Nova on AWS Bedrock
Smart sorting. Less waste. Better planet.

Built With

  • amazon-bedrock
  • amazon-nova-(us.amazon.nova-lite-v1:0)
  • amazon-web-services
  • bash
  • boto3
  • css3
  • css3-animations
  • fastapi
  • fetch-api
  • fingerprinting
  • html5
  • html5-canvas-api
  • javascript-es6+
  • loguru
  • mediadevices-api
  • osascript
  • pillow
  • pydantic
  • python-3.12
  • python-dotenv
  • sha-256
  • sqlite
  • uvicorn
  • vanilla-javascript
Share this project:

Updates