Real-Time Image Captioning System for Accessibility

Problem statement

Generating accurate, real-time image descriptions required balancing model complexity, latency, and user interaction constraints for accessibility use cases.

Architecture overview

Images were processed using distributed Spark jobs and TensorFlow-based object detection models. Generated captions were converted to speech output and integrated with haptic feedback for improved user interaction.

Technical decisions & tradeoffs

• Prioritized low-latency inference over model size.
• Used distributed processing to handle bursty workloads.
• Optimized interaction flow rather than raw accuracy metrics.

Lessons learned

Accessibility-focused systems demand reliability and predictability. User trust depends more on consistent behavior than peak accuracy.