The problem
An AI wearable needs all-day battery, a colour display, and an honest privacy story. Most ship cloud-only with a four-hour battery and a backhauled microphone.
EDGE AI · CASE STUDY
2023–26Brilliant Labs Halo — On-Device AI Wearable
Open-source AI glasses running on-device NPU inference, paired with a cloud agent with persistent memory.
Summary
Brilliant Labs Halo is an edge AI engagement running 2023–26. Open-source AI glasses running on-device NPU inference, paired with a cloud agent with persistent memory. Halo shipping with the developer community; full hardware and firmware stack open-source.
Facts
- Client
- Brilliant Labs Halo
- Sector
- Edge Ai
- Dates
- 2023–26
- Key result
- 14-hour battery · on-device NPU
- Site
- brilliant.xyz
- Engagement
- Discovery sprint → build alongside → hand-off
The architecture
Alif B1 (Cortex-M55 + NPU) in the frames, colour microOLED, bone-conduction audio. Cloud agent (Noa) for long-horizon tasks. Full stack — ZephyrOS, Lua runtime, schematics — open to developers.
- 01SiliconAlif B1 (Cortex-M55 with NPU) in the frames, sized so inference fits the accelerator's memory budget.
- 02Display and audiocolour microOLED with bone-conduction audio, no backhauled microphone.
- 03RuntimeZephyrOS with a Lua layer, so developers extend the device without a firmware build.
- 04Cloudthe Noa agent carries long-horizon tasks and persistent memory, off the critical path.
The results
Halo shipping with the developer community; full hardware and firmware stack open-source.
- Battery, mixed use
- 14 hoursAgainst roughly four hours for cloud-only wearables
- NPU inference
- On-deviceAudio is processed in the frames; the privacy story is structural
- Hardware and firmware
- Open stackSchematics, ZephyrOS build, and Lua runtime open to developers
“All-day battery and an honest privacy story are the same engineering decision: keep the inference on the device.”
Written by Binary AI Labs · Reviewed