Inspiration
Living in Ho, Ghana, I noticed a significant disconnect between local community clean-up efforts and our agricultural sector. We generate a lot of waste that ends up in landfills, yet local farmers are constantly in need of affordable, high-quality organic compost. I wanted to build a bridge between these two realities, but I realized that tracking community members' daily clean-up routes or exposing farmers' specific resource needs could compromise sensitive personal and economic data. AgricCycle AI was inspired by the idea that community waste is an untracked agricultural resource that can be routed to the right place safely, using zero-knowledge technology to protect everyone's privacy.
What it does
AgricCycle AI is a privacy-first web platform that empowers communities to manage waste intelligently. Users snap and upload photos of community waste, and our computer vision model automatically classifies the materials into organic compostables versus non-organic recyclables. Beyond simple classification, the platform features a localized resource mapping dashboard. Using the Midnight Network, community members can anonymously log waste hotspots and earn community incentives without ever exposing their personal identities, home addresses, or location histories on a public ledger. Local farmers can then securely access this aggregated, privacy-protected data to locate and collect compostable materials
How we built it
The core machine learning pipeline is built with Python, utilizing TensorFlow and Keras. I trained a Convolutional Neural Network (leveraging a fine-tuned ResNet50 architecture) to handle the image classification, leaning on pandas and scikit-learn for data processing.For the user interface, I deployed the application using Streamlit. I designed a custom UI with a "soft tech" aesthetic, incorporating deep red and burgundy accents to make the experience feel warm and inviting.To introduce the privacy layer, we integrated Midnight Network's capabilities. We utilized smart contracts to handle the anonymous logging of waste data and the secure distribution of reward tokens, ensuring that individual user data remains shielded while the aggregate maps remain highly useful for the farmers.
Challenges we ran into
Curating a dataset that accurately reflected the specific types of waste found in local community environments was a major hurdle, as pre-existing datasets lack localized context. Technically, the biggest challenge was bridging our Web2 machine learning stack with a Web3 privacy layer. Optimizing the deep learning model to run inference quickly within Streamlit while simultaneously handling secure, zero-knowledge smart contract interactions required a very careful balance of computational efficiency and system architecture.
Accomplishments that we're proud of
I am incredibly proud of successfully merging a deep learning computer vision model with zero-knowledge blockchain privacy. Moving from Jupyter Notebooks to a live web app is always satisfying, but proving that functional community tech can protect user data while still being beautifully and thoughtfully styled is a major win.
What we learned
This project pushed me to think beyond just model accuracy and focus on end-to-end deployment with data protection in mind. I learned a lot about bridging the gap between backend machine learning architectures, frontend user experience, and privacy-preserving cryptography specifically how to present complex, secure data in a way that is immediately useful to a non-technical end user like a local farmer.
What's next for AgricCycle AI
The next step is to expand our image dataset by crowd-sourcing photos from local clean-up initiatives to improve the model's accuracy on edge cases. I also plan to pitch the platform to local agricultural co-ops in Ho to pilot the privacy-preserving resource-mapping dashboard in a real-world setting, and eventually wrap the Streamlit app into a Progressive Web App (PWA) for better mobile accessibility in the field.
Built With
- github
- pandas
- python
- scikit-learn
- streamlit
- tensorflow
- vscode
