BloomShield

AI web app for real-time crop disease detection via a custom-trained YOLOv8m model (99.7% accuracy). Offers diagnoses, treatments, and community fine-tuning to cut farmer losses sustainably.

Comment

Inspiration

Crop diseases cause up to 30% yield loss globally, threatening food security and livelihoods, especially for small farmers. Inspired by this urgent problem and the vision of using ethical, authentic AI for social good, I set out to build a web app that empowers anyone to detect crop diseases instantly removing barriers of cost, access, and technical complexity.

What it does

BloomShield is an AI-powered web app that enables users to upload a plant image and receive real-time, highly accurate disease diagnosis, along with actionable treatment and prevention advice. The app uses a YOLOv8m model trained from scratch on 87,000+ labeled images spanning 38 crop health classes. It also invites the community to contribute new labeled images, making the model smarter over time.

How we built it

  • Model: Trained YOLOv8m, using PyTorch and the Ultralytics framework, on the New Plant Diseases Dataset with 7 epochs and advanced augmentation, achieving over 99% accuracy.
  • Backend: Flask API in Python serves predictions and handles user/community uploads.
  • Frontend: Built with minimalist Swiss-style HTML/CSS/JS, prioritizing accessibility and mobile-first design.
  • Database: SQLite stores community-contributed images and metadata.
  • Deployment: Hosted on Render.com using Gunicorn for production reliability.
  • Open Source: Full code and training workflows are public on GitHub.

Challenges we ran into

  • Model Integration: Initial attempts with YOLOv9 led to compatibility issues and integration errors—resolved by switching to YOLOv8m and retraining from scratch.
  • Large Dataset Handling: Managing and preprocessing 87K+ images required efficient scripting and batch operations within limited resources.
  • Deadline Constraints: Building a robust ML web app, while ensuring transparency, reproducibility, and a clean, accessible UI all under hackathon time pressure.

Accomplishments that we're proud of

  • Achieved 99.7% Top-1 classification accuracy, demonstrating authentic, high-quality ML results.
  • Designed a minimalist, user-friendly UI that works beautifully on mobile and desktop.
  • Implemented a true community-driven training loop, enabling ongoing model improvement.
  • Fully open-sourced every component code, training logs, dataset references, and UI—for maximum reproducibility and social impact.

What we learned

  • Authentic ML from data curation to architecture selection and evaluation matters not just for competition rules but for real-world, reliable AI.
  • Swiss-style minimal frontend design leads to better accessibility and faster adoption among non-technical users.
  • Iterative development, fast bug-fixing, and committing every step to version control create demonstrably more robust and trustworthy software.

What's next for BloomShield

  • Expand to support multi-language interfaces and smallholder-friendly offline workflows.
  • Add explainability features so farmers can see why a particular disease was diagnosed.
  • Build partnerships with agri-extension networks to deploy BloomShield at scale.
  • Continue crowdsourcing new data to keep the model up to date—and open—to benefit global agriculture.

Built With

Share this project:

Updates