EcoSort Edge

Inspiration

Modern consumer packaging rarely consists of a single material. Everyday items—like coffee cups lined with polyethylene, blister packs fusing plastic and aluminum, or shipping cartons layered with adhesive tapes—confuse conventional waste management pipelines. When multi-material packaging enters standard recycling streams, entire batches become contaminated, driving down recycling yields and overburdening municipal landfills. We built EcoSort Edge to bridge visual inspection with multi-layer material decomposition right at the source.

What it does

EcoSort Edge is an intelligent multimodal vision pipeline designed to inspect, isolate, and route complex consumer packaging:

• Visual Material Isolation: Ingests live image feeds of discarded items and decomposes them into distinct structural sub-components (e.g., separating the cardboard sleeve, plastic lid, and waxed paper body).
• Multi-Modal Property Reasoning: Analyzes spatial geometry, surface texture, printed resin codes (RIC 1–7), and material density indicators simultaneously.
• Granular Recyclability Scoring: Computes recyclability indices per sub-component and outputs specific disposal instructions (e.g., "Separate aluminum foil backing before composting cardboard tray").
• Edge-Optimized Diagnostics: Delivers real-time local classification and structured telemetry output without relying on high-latency cloud backends.

How we built it

• Application Core & Interface: Developed in Python using Streamlit for an interactive, low-latency live inspection console.
• Multimodal Perception Layer: Connected vision and multimodal reasoning models to extract element hierarchies, component boundaries, and material compositions in a single pass.
• Prompt & Schema Engineering: Formatted model responses using strict JSON schemas to output reliable key-value pairs (material type, polymer grade, contamination risk, sorting destination).
• Environment & Tools: Prototyped and refined within Visual Studio Code, tracked via Git and GitHub.

Challenges we faced

• Visual Ambiguity in Deformed Packaging: Crushed, crumpled, or wet items often obscure resin identification codes and distort standard silhouettes, requiring extensive prompt conditioning on edge textures and reflection patterns.
• Latency vs. Accuracy: Balancing dense visual scene parsing with fast response times for real-time edge sorting scenarios.
• Composite Layering: Distinguishing between laminated composites (which cannot be separated mechanically) and multi-part assemblies (which can be detached by consumers).

Accomplishments that we're proud of

• Successfully structuring complex packaging into modular, actionable constituent parts rather than settling for generic top-level labels like "plastic" or "paper".
• Designing a clean, functional web interface capable of handling image ingestion, live parsing, and standardized disposal telemetry seamlessly.

What we learned

• Effective multimodal pipelines require precise architectural framing: asking the model to visually isolate components before scoring recyclability dramatically reduces hallucinated classifications.
• How to design modular interfaces that bridge complex reasoning backends with intuitive end-user directives.

What's next for EcoSort Edge

• Edge Hardware Integration: Deploying lightweight quantized models onto edge camera setups (e.g., Raspberry Pi or Jetson platforms) for physical bin attachment.
• Optical Sorter Actuation: Connecting output signals to automated pneumatic diverters and robotic sorting mechanisms.
• Extended Materials Database: Expanding support for industrial polymers, e-waste sub-assemblies, and bio-plastics.

Built With

• computer
• git
• github
• json
• multimodal
• python
• streamlit
• vision