The Problem
When families arrive at USAG Stuttgart in Germany, they face an unexpected culture shock: five colored recycling bins, strict contamination rules, and instructions entirely in German. Over 70% of newcomers reported confusion about proper recycling in our initial surveys. Contamination isn't just embarrassing — it can result in fines and violations of the Status of Forces Agreement (SOFA) governing U.S. military communities abroad.
I wanted to fix this with AI.
What I Built
GelbIT (German for yellow — the recycling bin color that trips people up most) is a web app that lets you photograph a piece of trash and instantly get precise, location-aware recycling instructions in your language.
Three AI systems work together:
1. Fine-tuned Computer Vision A customized Gemma model classifies not just material type but contamination and condition. Instead of "metal can," it outputs "metal can with sauce residue" — which changes the recycling guidance entirely. I assembled a dataset of 2,527 labeled images specific to common household items in our community.
2. Agentic Research System Rather than a static database (which goes stale the moment local rules change), I built an agentic system using Python, Selenium, and Ollama. When given a classified item, it autonomously formulates search queries, scrapes official municipal websites and the Stuttgart Amt für Umweltschutz, cross-references multiple sources, and synthesizes current, hyper-local guidance. It even generates follow-up queries automatically — detecting sauce on a can triggers additional research on whether Stuttgart's facilities accept contaminated metals.
3. Multi-Language NLP Translation and natural language generation ensure the app works for English, German, Spanish, and other languages common in our diverse community.
How It Works
- User uploads or captures a photo
- Gemma classifies the item (with contamination/condition context)
- Agentic system live-scrapes Stuttgart's waste management authority for current rules
- App returns: item classification, specific bin instructions, any prep steps (rinsing etc.), and a map of nearby recycling centers
The whole flow takes seconds.
Engineering Challenges
Running on consumer hardware. Most AI pipelines require cloud infrastructure or expensive GPUs. I deliberately optimized to run on my personal laptop — careful model optimization, query caching, and smart resource management. Any community organization could deploy this.
Accuracy at the edge. Recycling rules vary by municipality and change over time. My multi-source verification approach has the agent flag discrepancies and present confidence scores. When the AI is uncertain, it says so and links the user to official sources rather than guessing.
Web scraping at scale. Dynamic JS-rendered content, varying site structures, and respectful rate limiting all required thoughtful engineering — Selenium browser automation, adherence to robots.txt, and fallback sources when primaries were inaccessible.
Results
- 94% accuracy across 100 diverse test items against official Stadt Stuttgart guidelines
- 85% self-reported increase in recycling participation from beta users
- Zero contamination notices reported by beta users during two-week real-world deployment
- Vice President of our school, after first use: "No plastic bags?! Oops!!!! I'm doing it wrong. You already helped me!!"
- Facility managers have requested deployment in common areas for bulk waste guidance
- Exploring expansion to other European military installations
What I Learned
The biggest lesson: responsible AI means designing for failure gracefully. Our image classification struggled with transparent plastics and poor lighting. Rather than hiding that, I built it into the UX — confidence thresholds, fallback manual selection, and clear prompts like "photograph items in good lighting." Trust comes from transparency, not from pretending the AI is infallible.
I also learned that the most impactful AI applications solve boring, everyday problems. Recycling isn't glamorous. But eliminating genuine confusion for thousands of people — and doing it with AI that actually works — matters more than a flashy demo.
Built with: Python, Gemma (fine-tuned), Selenium, Ollama, LM Studio, geolocation API, multi-language NLP. Team: Maximilian Pezzullo & William Hargrove. Supervising Adult: Matthew Snoeyink.