Embodied AI · edge + hosted

WearabLLM

A wearable voice interface that combines onboard audio, expressive physical feedback, and a protected hosted-agent path—so the interaction is more than another chat window.

Back to physical computing

The question

What changes when an AI companion’s primary output is not another chat window, but a physical and expressive presence? WearabLLM treats the device’s audio, color, and timing as part of the response itself.

What I built

I built and flashed firmware for a Waveshare ESP32-S3 audio board with microphone capture, speaker output, an RGB response language, BOOT/PTT activation, and a local Hi ESP wake path. The normal design sends audio over Wi-Fi to a protected hosted service; a local bridge remains available for development and richer memory experiments.

Physical initiation: BOOT/PTT or local wake phrase Onboard interaction: microphone capture, RGB state, speaker output Protected hosted path: transcription, language response, and text-to-speech Optional private continuity: bounded session context and transcript logging

Engineering through the abstraction axis

The project moves deliberately between board-level constraints and service-level behavior: I2S audio capture, firmware configuration, network requests, response normalization, TTS playback, and a carefully limited memory/logging path. That makes it a useful example of how system-design decisions show up directly in an embodied interface.

What is verified

The current hardware project has documented local builds and flashes, observed board boot, wake/PTT capture, non-silent microphone recording, physical speaker tones, valid TTS playback, and session-context recall through the bridge. Firmware, bridge, app, and protocol checks are also maintained separately.

Validation boundary

A single end-to-end physical session that combines every hosted step—wake or PTT, microphone capture, remote request, LED response, audible TTS, and transcript persistence—still needs a final focused confirmation. This public page intentionally does not expose device-control endpoints, private logs, or credentials.