← All work

ScanBench

Can a phone run a vision-language model fast enough to describe the world in real time?

July 2026

  • On-device ML
  • Android
  • Kotlin
  • Accessibility

A feasibility probe for a scanning assistant for blind and low-vision users. The app sends one fixed prompt with a camera frame every N seconds, fully on-device with Gemma 4 E2B. It breaks down where the milliseconds go: time to first token (when you could start speaking), prefill cost of the image, decode throughput, JPEG encode and end-to-end capture-to-result. It also shows a rolling headroom readout against the interval, so the question goes from “does it run” to “does it keep up on this phone”.