CodeNetra AI: Inclusive Vision & Dev Suite

A Dual-Mode Super App powered by Gemini 3 Flash. It provides "Digital Eyes" for the blind and a "Coding Brain" for developers in one inclusive suite.

Comment

💡 Inspiration

"Netra" means Eye in Sanskrit. I asked myself a single question: Can one AI architecture solve two completely different human problems?

  1. The Darkness: Millions of visually impaired people need "Digital Eyes" to navigate safely.
  2. The Burnout: Developers need a "Digital Brain" to understand complex legacy codebases instantly.

CodeNetra-AI is the answer. It is not just an app; it is a Dual-Mode Multimodal Bridge powered by the sheer raw power of Gemini 3.0.

🚀 What it does

CodeNetra is a Super-App with two distinct personalities:

👁️ Mode 1: Netra Vision (Accessibility Suite)

  • The "Jarvis" Loop: A completely hands-free, continuous voice loop. You speak, it answers, it listens again. No buttons.
  • Live Hazard Defense: Detects cars, pits, and obstacles in real-time and screams "Caution!" before you hit them.
  • The "Aura": Ask "Who made you?" and experience a unique AI identity that defines its own existence.

💻 Mode 2: Developer God-Mode (Productivity Suite)

  • Talk to your Repo (ZIP Intelligence): Upload an entire project .zip. CodeNetra reads the whole architecture and explains how the login flow works.
  • UI to Code: Show it a screenshot, and it writes the clean Flutter/React code instantly.
  • AI Interviewer: It takes your technical mock interview to prep you for FAANG.

⚙️ How we built it

We pushed Gemini 3.0 Flash to its limits using Flutter.

  • The Brain: Used gemini-3-flash-preview for ultra-low latency. Speed was our #1 priority for blind safety.
  • The Ears: Built a custom Voice Loop Engine that manages state between Listening -> Thinking -> Speaking to create a seamless conversation flow (1.5s latency).
  • The Context: Leveraged Gemini's Long Context Window to ingest massive ZIP files without crashing.

🚧 Challenges we ran into

  • The "Infinite Loop" Bug: Initially, the AI would listen to its own voice and reply to itself forever. I fixed this with a smart SpeechStop logic.
  • Token Limits: Uploading large codebases hit the limit fast. I wrote a filter to strip out assets/build folders before sending data to Gemini.
  • Git Security: My API keys got exposed in history. Learned the hard way how to use .gitignore and git rm --cached to secure the project.

🏅 Accomplishments that I'm proud of

  • The "Killer Aura": Successfully giving the AI a personality. It feels alive, not robotic.
  • Latency Optimization: Bringing the Vision-to-Voice response time down to under 2 seconds.
  • Solo Development: Building a Dual-Mode Enterprise-level app alone.

🧠 What I learned

Multimodality is the future. Text is not enough. The ability of Gemini to "See" (Vision) and "Hear" (Voice) simultaneously is what makes inclusive apps possible.

🔮 What's next for CodeNetra-AI

  • Smart Glasses: Moving "Netra Mode" from phone to wearable glasses.
  • Offline Mode: Using Gemini Nano for on-device object detection without internet.

Built With

Share this project:

Updates