Inspiration
Navigating everyday environments is heavily visual, which creates constant friction for visually impaired individuals. While many assistive tools exist, they often depend on smartphones, screens, or high-latency cloud interactions that break the immediacy needed for real-world navigation.
We were inspired by the idea that assistive intelligence should live as close to the user as possible—responding quickly, respecting privacy, and functioning even under limited connectivity. Edge computing offers a natural foundation for accessibility-first systems.
What it does
EdgeSight Voice Assistant for the Visually Impaired is an edge-controlled vision system that helps users understand their surroundings through spoken feedback.
The system:
- Captures images locally using an on-device camera
- Sends visual data for semantic understanding only when needed
- Converts scene descriptions into real-time voice output
- Operates hands-free to support accessibility
- Exposes latency and processing stages to maintain transparency
Its goal is to provide immediate, understandable context about nearby objects and environments without requiring visual interaction.
How we built it
EdgeSight is built using a hybrid edge architecture that prioritizes local control while selectively leveraging cloud intelligence.
Edge components
- Local camera capture and trigger logic
- On-device orchestration and latency measurement
- Offline text-to-speech for voice feedback
Cloud-assisted components
- Vision-based scene understanding using an external AI model
- Natural language generation optimized for accessibility
We deliberately separated edge responsibilities from cloud processing to ensure low-latency interaction, privacy awareness, and a clear path toward future on-device inference.
Challenges we ran into
- Balancing edge constraints with model capability: Vision models are computationally expensive, requiring careful decisions about what remains local versus remote.
- Latency sensitivity: Accessibility systems must respond quickly, so we instrumented and monitored every stage of the pipeline.
- Avoiding misleading claims: It was critical to communicate architectural limits honestly rather than overstating “fully on-device” capabilities.
- Usability without screens: Designing a workflow that is entirely voice-driven required repeated simplification.
Accomplishments that we're proud of
- Delivered a functional assistive system aligned with real edge constraints
- Built a hands-free, voice-first interaction model
- Demonstrated an honest and measurable edge-to-cloud boundary
- Created a system that prioritizes accessibility and privacy
- Completed a working demo within a short build window
What we learned
- Edge systems are defined more by control and latency than by isolation from the cloud
- Accessibility requirements expose architectural weaknesses quickly
- Hybrid designs are often the most practical path toward real-world deployment
- Measuring and explaining system behavior builds more trust than claiming novelty
- Simplicity significantly improves usability in assistive technology
What's next for EdgeSight Voice Assistant for the Visually Impaired
- Replace cloud vision with lightweight on-device models as hardware permits
- Integrate continuous audio input for conversational interaction
- Expand object and context recognition for indoor and outdoor environments
- Optimize for deployment on dedicated edge hardware such as smart glasses
- Conduct user testing with visually impaired communities to guide iteration
Log in or sign up for Devpost to join the conversation.