Zardon Vision

An AI-powered visual assistant that uses object detection, OCR, and voice feedback to help visually impaired users navigate, read, and interact with the world independently.

Comment

Inspiration

One afternoon at the library, we met a young boy struggling to read because he was visually impaired and had forgotten his glasses. Seeing how something so small could create such a big barrier inspired us to build Zardon Vision — an AI-powered mobile assistant designed to increase independence and accessibility for visually impaired individuals.

What it does

Zardon Vision transforms a smartphone into a real-time visual assistant.

  • 📷 Object & human detection using computer vision
  • 📖 Text-to-speech (OCR) to read books, signs, and documents aloud
  • 🧭 Depth awareness to better understand surroundings
  • 🗣️ Siri shortcut integration for hands-free requests and emergency contact access

The app converts visual information into audio, allowing users to better navigate and understand their environment.

How we built it

Frontend: SwiftUI + UIKit
Backend: Python + Flask
AI/ML: PyTorch, YOLOv2, Create ML
Text Recognition: Apple VisionKit & AVFoundation
Website: HTML + CSS

Images are captured on-device and sent to our Flask server, where custom-trained PyTorch models classify objects and return results to be read aloud. For text, Apple’s Vision APIs detect and extract written content, which is converted into speech instantly.

Challenges we ran into

  • Overfitting and low accuracy in facial/object recognition
  • No fallback (“catch case”) for uncertain classifications
  • Flask server connection issues on restricted WiFi
  • SwiftUI limitations that required UIKit workarounds
  • Designing a UI truly accessible for visually impaired users

Accomplishments that we're proud of

  • Successfully training and deploying custom YOLOv2 models
  • Integrating real-time OCR with audio output
  • Building Siri shortcut voice activation in under 10 seconds
  • Creating a clean, high-contrast, accessibility-first interface
  • Delivering a fully functional end-to-end AI-powered system

What we learned

  • Model training requires diverse datasets to avoid overfitting
  • Accessibility design is more than aesthetics — it’s usability and clarity
  • Backend reliability is just as important as model performance
  • Building AI for real-world users demands iteration and testing

What's next for Zardon Vision

  • Improve model accuracy with larger datasets
  • Add fallback classifications for uncertain predictions
  • Enable offline inference
  • Expand multilingual OCR support
  • Enhance depth-based navigation features

Zardon Vision is just the beginning of our mission to make AI-driven accessibility tools more powerful, reliable, and widely available.

Built With

Share this project:

Updates