Why on-device AI is the right foundation for assistive technology
Cloud AI fails exactly when assistive tech users need it most. Here's why EchoSense runs every model on-device — for privacy, latency, cost, and reliability.
Most AI products send your words to a data center and stream the answer back. That architecture is acceptable for a chatbot. It is not acceptable for a device a blind person depends on to cross a street, or an app an autistic teenager confides in daily.
Four reasons assistive AI must run on-device
- Reliability — a navigation aid that stops working in a metro station, an elevator, or a village with weak coverage is worse than no aid at all. Echo 1's obstacle detection, voice assistant and navigation all work with the SIM removed.
- Privacy — conversations with a companion app include health details, fears, and family matters. The only way to guarantee they stay private is for them never to leave the device.
- Latency — an obstacle warning that arrives 800 ms late is a collision. On-device inference on the Hailo-10H delivers ~17 ms per frame.
- Cost — cloud inference means subscriptions. Assistive technology priced as a monthly fee excludes exactly the users who need it most.
What 'on-device' means in EchoSense, concretely
On the Echo 1 wearable, YOLOv8n runs on a Hailo AI accelerator, faster-whisper transcribes speech, SmolVLM2 describes scenes, and Tesseract reads text — all on a Raspberry Pi 5. In the Companion app, Gemma 4 runs on the iPhone itself via Apple MLX, handling text, voice, and images.
The cloud still exists in our architecture, but it is demoted: it handles accounts, caregiver linking, and sync. It never sees the content of a conversation. That is a design constraint, not a settings toggle.
Questions about EchoSense, pilots, or partnerships?
Talk to the team