ManuAI
A voice helper for factory workers — say what's wrong, get a cited procedure or a safety warning, all on a Mac with wifi off. An iOS app sends Ray-Ban mic audio into a local Whisper → Moss → Qwen pipeline over WiFi or USB.
Edge runtime
mlx-whisper transcribes the utterance, Moss retrieves approved SOP chunks in single-digit milliseconds from a local index, and Qwen2.5 composes a forced-JSON cite-or-refuse answer. Kokoro TTS speaks it back; every consumer renders the same screen_state contract — answer, citations, safety flags, and escalation.
iOS glasses bridge
The ManuAI iOS app (Meta Wearables DAT SDK) captures Ray-Ban HFP microphone audio and streams Float32 48 kHz PCM to glasses_bridge.py on the Mac. Hands-free mode skips the camera; an air-gapped usbmux path inverts roles so the phone is a TCP server and the Mac connects over USB with Wi-Fi and cellular off.
Retrieval and corpus
Unsiloed ingests SOP PDFs into page-aware chunks with error codes and safety metadata. A background context swarm prefetches related LOTO and incident context while the foreground answers. Operator chat logs provide a secondary corroboration index against how similar faults were actually resolved on the floor.
Delivery modes
Three entry points share one brain: offline_demo.py (laptop mic, WebRTC-free wifi-off headline), glasses_bridge.py (Ray-Ban input, laptop speaker + live SOP screen), and operator.html with LiveKit push-to-talk for the wifi-on browser path. The operator console also renders a live Ray-Ban MJPEG feed and a per-turn context graph.