Building a wearable on Raspberry Pi 5 + Hailo: lessons from Echo 1

Echo 1 runs on a Raspberry Pi 5 (16 GB) with a Hailo AI HAT+ delivering 40 TOPS, on Debian 13. Getting from a bag of parts to a device with 121 passing hardware tests taught us more than any course. Here's what transfers.

Bench-test every sensor before integration

Every component — camera, LiDAR, GPS, microphone, speaker, button — was validated alone, with a scripted bring-up, before meeting the others. When something breaks in an integrated system, you want to already know each part works; otherwise every bug has six suspects.

Be willing to un-choose hardware

The original design included an MPU-6050 IMU and a SIM808 GSM module. Both were dropped — each via a documented architecture decision record — when validation showed they didn't earn their complexity. Killing a component you already wired is painful; shipping flaky hardware to a blind user is worse.

The details that cost us days

• ALSA device indices reorder across boots — address audio devices by name, never by index.
• The INMP441 I2S mic delivers real audio on the left channel; the right is silence. Undocumented, discovered the hard way.
• UART budgeting — LiDAR at 115200 and GPS at 9600 need their own ports and their own owning process; sharing ends in corrupted frames.
• GPIO polling at 50 Hz with debouncing beat interrupt-driven button handling for reliability on a device that gets bumped all day.
• No hardcoded constants — every pin, port, and threshold lives in one config file with env-var overrides, so hardware revisions don't mean code hunts.

The Hailo changes what's possible

The Pi 5 alone runs small models slowly and hot. The Hailo-10H runs YOLOv8n at ~17 ms per frame, freeing the CPU for whisper transcription, VLM inference, and OCR. Commodity board plus commodity accelerator is the price-performance combination that makes an affordable wearable real.