Gdzie to wyrzucić? 🗑️ [ENG: Where to throw it?]

A Polish recycling assistant powered by Gemma 4

Take a photo of any item — even contaminated or unusual ones — and find out which bin it belongs in. Built for the dev.to Gemma 4 Challenge, May 2026.

🔗 Live demo: gdzie-wyrzucic.vercel.app

📝 Write-up: dev.to article

---

The problem

Recycling rules in Poland confuse even people who care. The hard cases aren't "plastic bottle" or "newspaper" — they're:

• An empty hair dye bottle stained with dried dye
• A pizza box with grease on the bottom
• A pet food can with dried residue
• Old electronics: where exactly do they go?
• The new deposit system (kaucja, launching January 2026) — which bottles qualify?
• The new purple textile bins (mandatory since January 2025)

Existing apps in Poland are searchable databases — they require users to know what an item is called in Polish, and they often miss edge cases. This app uses Gemma 4's vision capabilities to look at the item and reason about it, including its condition.

How it works

1. User takes a photo of an item with their phone camera
2. The app sends the photo to Gemma 4 26B Mixture-of-Experts via Google AI Studio API, along with a detailed system prompt encoding Polish recycling rules
3. Gemma returns structured JSON: bin category, preparation steps, explanation, edge-case notes
4. The app displays the result with the corresponding bin color

Categories supported

The app classifies items into twelve disposal categories matching Poland's current waste system:

Category	Bin color	Examples
plastik_metal	🟡 Yellow	Plastic bottles, cans, Tetra Pak
papier	🔵 Blue	Paper, cardboard
szklo	🟢 Green	Glass jars, bottles
bio	🟤 Brown	Food scraps, garden waste
zmieszane	⚫ Grey	Mixed waste — contaminated or composite items
tekstylia	🟣 Purple	Clothes, shoes (separate collection mandatory since Jan 2025)
elektroodpady	🔴 Red	Small electronics, batteries
pszok	—	Large/hazardous items (municipal drop-off)
kaucja	—	Deposit-return bottles & cans (launching January 2026)
niewyrazne	—	Photo too unclear to classify — prompts user to retake
niewaste	—	Photo doesn't show a waste item
apteka	—	Expired or unused medications — pharmacy drop-off

Why Gemma 4 26B MoE

Gemma 4 comes in four variants. I chose 26B MoE (gemma-4-26b-a4b-it) for these reasons:

• Production-quality reasoning at edge-class compute. The 26B MoE architecture activates only ~4B parameters per token, so it reasons at 26B-class quality while running at roughly E4B cost. Hard recycling edge cases — composite packaging, contaminated containers, blister packs photographed from the foil side — benefit from the larger knowledge base.
• Better accuracy on ambiguous items. Classifying items correctly matters more than classification speed here. The 26B MoE handles material composition, contamination tradeoffs, and subtle visual cues more reliably than smaller variants.
• Free hosting via Google AI Studio enables 24/7 availability. No local infrastructure, no tunnel, no cold-start latency — the demo runs as a standard Vercel serverless function calling a managed API.
• The on-device story still holds. During development I tested E4B locally via Ollama to validate that the smaller variant can handle the task on-device. The deployment choice (cloud-hosted 26B) is separate from the capability story: E4B works on-device too, making a future MediaPipe or local-Ollama deployment viable without rewriting the system prompt.

Tech stack

• Next.js 16 (App Router, TypeScript)
• Tailwind CSS v4 for styling
• Gemma 4 26B MoE (gemma-4-26b-a4b-it) via Google AI Studio API
• Sharp for server-side image resizing before inference
• Vercel for deployment

Running locally

You need a Google AI Studio API key (free tier available).

# Clone the repo
git clone https://github.com/YOUR-USERNAME/recycling-app.git
cd recycling-app

# Install dependencies
npm install

# Set your Google AI Studio credentials
echo "GOOGLE_API_KEY=your_key_here" >> .env.local
echo "GOOGLE_MODEL=gemma-4-26b-a4b-it" >> .env.local

# Run the dev server
npm run dev

Open http://localhost:3000 in your browser.

Deploying to Vercel? Add GOOGLE_API_KEY and GOOGLE_MODEL as environment variables in your Vercel project settings.

Project status

This is a hackathon submission built solo in ~10 days. It's a working demo, not production software. Known limitations:

• Recycling rules vary by gmina (Polish municipality); the app uses general national guidelines
• Not all bin types (especially purple textile bins) are available in every neighborhood yet
• The kaucja symbol recognition depends on photo quality
• Confidence calibration is rough — Gemma may be confidently wrong on unusual items

Future ideas: location-aware gmina-specific rules, integration with PSZOK location databases, on-device deployment via MediaPipe, multi-language support.

About

Built by Klaudia Grzondziel for the dev.to Gemma 4 Challenge.

I'm a technical writer, not a developer — this is my first web app. AI coding assistance via Claude Code helped me bridge the gap from idea to working code. The product design, system prompt engineering, and Polish recycling knowledge are mine; the boilerplate code is mostly AI-assisted.

License

MIT — see LICENSE.