WasteSnap

Inspiration

Recycling systems fail primarily at the point of disposal. Studies indicate that up to 25% of collected recyclable material ends up in
landfills due to contamination caused by mis-sorting. People genuinely want to recycle correctly, but packaging materials are
increasingly complex, local municipal guidelines differ, and resin identification codes on plastics are hard to read.

We built WasteSnap to eliminate guesswork at the bin. By combining real-time computer vision with geospatial routing, WasteSnap turns any
smartphone camera into an instant sorting guide that not only identifies what an item is and how to dispose of it, but also directs
users to the nearest verified drop-off point.

What It Does

WasteSnap is a progressive web application (PWA) designed for zero-friction waste triage:

1. Snap & Classify: Point the device camera at any discarded item or packaging. Vision AI identifies the item, primary material
composition, disposal category (recyclable, compostable, landfill, hazardous, or e-waste), confidence score, and clear
handling instructions. Non-waste items are rejected automatically.
2. Geospatial Facility Routing: When location access is enabled, WasteSnap queries OpenStreetMap geospatial data via the Overpass API
within a 5 km radius, calculating Haversine distances to the nearest drop-off or recycling facility matching that specific waste
category.
3. Footprint Dashboard & Streak Tracking: Every scan persists locally in localStorage, calculating the user's diversion rate (% diverted
from landfill), logging consecutive daily streaks, and rendering a breakdown chart.
4. Export & Shareable Verdict Cards: Users can export their full scan log to CSV or generate a clean summary image card via canvas to
share on social channels.
5. Bilingual Accessibility: Supports both English and Bahasa Indonesia with dynamically localized prompts so classification instructions
match the user's language.

The Waste Problem Statement & Impact Measurement

Targeted Problem: Municipal recycling contamination and low capture rates of hazardous/electronic waste caused by lack of actionable
information at the exact moment of disposal.

Measurement Methodology:
We measure real-world impact across three objective metrics:

1. Classification Accuracy Rate: Targeting ≥ 90% correct categorization across standard municipal solid waste categories (benchmarked
against municipal sorting datasets).
2. Diversion Rate Impact: Track the percentage of scanned items successfully categorized as recyclable or compostable versus landfill:

  Diversion Rate = (N_(recyclable) + N_(compostable))/Nₜₒₜₐₗ × 100%

3. Habitual Retention: Tracking 7-day and 30-day repeat usage via client-side streak logs to measure whether users build consistent
sorting habits over time.

How We Built It

• Frontend & PWA: Built with vanilla HTML5, modern CSS, and JavaScript. Zero client-side framework bloat ensures instant loading and full
compatibility on low-end mobile devices without app store installation.
• Vision AI Inference: Powered by Groq's high-speed LPU inference engine running the qwen/qwen3.8-27b multimodal vision model, producing
structured JSON outputs with low latency (<1.5s).
• Geospatial Intelligence: Browser Geolocation API coupled with OpenStreetMap Overpass Over-The-Air queries using node tagging
(amenity=recycling, recycling_type=centre, recycling:small_appliances=yes). Distance is computed locally via the Haversine formula:

  d = 2R arcsin(√{sin²(Δφ/2) + cos(φ₁)cos(φ₂)sin²(Δλ/2)})
• Geospatial Intelligence: Browser Geolocation API coupled with OpenStreetMap Overpass Over-The-Air queries using node tagging
(amenity=recycling, recycling_type=centre, recycling:small_appliances=yes). Distance is computed locally via the Haversine formula:

  d = 2R arcsin(√{sin²(Δφ/2) + cos(φ₁)cos(φ₂)sin²(Δλ/2)})

• Hosting & Infrastructure: Deployed on Vercel edge infrastructure with continuous deployment from GitHub.                            

Challenges We Faced                                                                                                                   

1. Model Latency vs. Accuracy: Early prototypes tested heavier vision models that had 6–8 second latencies, which breaks the real-time   
utility of standing at a trash bin. Migrating to Groq-hosted vision inference reduced response latency to under 1.5 seconds.
2. Handling Unstructured Packaging: Household items often have multi-layered materials (e.g., coffee cups with plastic lining, blister
packs). Fine-tuning the prompt constraints ensured the model provides explicit handling instructions rather than generic labels.
3. Zero-Dependency Mobile UX: Ensuring clean camera hardware access across mobile browsers without mirrored viewfinder artifacts or heavy
frontend libraries required careful HTML5 canvas and video stream management.                                                          

What We Learned                                                                                                                     

• Real-world waste reduction tools must prioritize speed and zero-friction onboarding over complexity; requiring user accounts or large
app downloads drastically drops conversion.
• Open geospatial data (OpenStreetMap) provides an accessible, global foundation for public recycling infrastructure routing without
proprietary map API costs.                                                                     

What's Next for WasteSnap                                                                                                              

• Barcode / UPC Fallback: Integrate UPC database lookup for packaged goods with damaged labels.
• Municipal Geo-fencing: Integrate municipal-specific rule sets (e.g., distinguishing dual-stream vs. single-stream cities).
• Offline Edge Inference: Compile a quantized lightweight vision model into WebAssembly / TensorFlow.js for scanning with zero internet    
connection.

Built With

• groq
• html5
• javascript