Software Stack
| Layer | Choice | Rationale |
|---|---|---|
| Client SDK |
ao_reach (Dart ^3.5) |
Session overlays + reverse MCP tunnel, native to AO |
| Client shell | Dart bridge + Chromium kiosk | Bridge speaks REACH; kiosk renders; local WS between them |
| Avatar (Phase 1) | Live SVG starburst in Chromium | State, mic level, speech amplitude; no WebGL required |
| Avatar (planned) | TalkingHead.js + GLB | Lip-sync path reserved; GLB UAT still open |
| Vision |
CodeProject.AI Server (:32168) |
Already deployed, CUDA-backed |
| Wake word | openWakeWord | Free, local, custom words from synthetic audio |
| VAD | Energy VAD (Silero optional) | End-of-speech with hysteresis |
| STT | AO speech sidecar or local faster-whisper | Prefer Ada when hello.speech
|
| TTS | AO speech sidecar or local Piper/sherpa | OpenAI-compatible /v1/audio/speech
|
| Orchestration | agentic-orchestration ≥ v1.28 for speech advertise | Planner, agents, MCP, KB |
| # | Decision | Chosen | Rejected |
|---|---|---|---|
| 1 | Vision | CodeProject.AI | face-api.js, MediaPipe, custom InsightFace |
| 2 | Transport to AO | AO Reach | Plain REST /voice/query
|
| 3 | Client shell | Dart + Chromium | Pure Flutter / Electron |
| 4 | Inference split | Ada for AO/CPAI/speech; Pi for I/O | All-on-Pi forever |
| 5 | Wake word | openWakeWord | Porcupine, Precise |
| 6 | Duplex | Half (Phase 1) | Full duplex + AEC |
| 7 | Identity | Face → session headers | Prompt-prefixed context |
| 8 | Voice MCP set | Fast MCPs only | All MCPs (search/arxiv blow 15s) |
Long-form: ADR Index.
See Baselines for pinned observations (CPAI 2.9.x, AO 1.27.4 verified, speech path targeting 1.28+, Chromium on Pi, Python 3.11/3.12 for STT venv).