Integrated demo and architecture memo

Day 6: Integrated demo and architecture memo

今日目標

把前五天的知識變成可以展示的 evidence。

Demo scope

Demo 可以是 semi-real-time，不必假裝 production-ready。重要的是量測、界線與可重現。

Flow：

audio file or microphone
-> VAD / ASR
-> optional diarization label
-> PII redaction
-> RAG retrieve
-> agent feedback
-> output guardrail
-> TTS or text response
-> audit log

畫面或 README 必須列：

hardware
model names
runtime
input audio condition
ASR latency
LLM first token latency
TTS first audio latency
end-to-end latency
known limitations

Architecture memo template

標題：

Enterprise Voice Agent Gateway v0 Architecture Proposal

第一段：capability

本架構支援多種企業 AI Coach 任務。共同層負責 audio intake、model routing、
identity、policy、PII、tool permission、memory scope、audit、evaluation、
red teaming；adapter 層負責不同任務的 taxonomy、RAG corpus、output schema、
policy rules、tool permissions、evaluator。

第二段：layers

Audio layer:
  VAD / ASR / diarization / hotword correction / TTS

Knowledge layer:
  parsing / metadata / embedding / vector DB / reranker / citation

Agent layer:
  orchestrator / tool registry / memory / approval queue

Governance layer:
  identity / RBAC / ABAC / PII / policy / audit / red teaming

Deployment layer:
  Docker / K8s / vLLM / observability / rollback

第三段：validation

Functional:
  完成指定 coaching / report / retrieval task

Quality:
  ASR WER/CER, keyword accuracy, RAG hit@k/MRR, answer faithfulness

Latency:
  p50/p95 by component and end-to-end

Security:
  OWASP/NIST mapped red-team pass rate, PII leakage tests

Ops:
  health checks, logs, GPU memory, rollback, acceptance checklist

今日產出

建立：

demo-script.md
architecture-memo.md
known-limitations.md
acceptance-checklist.md

Known limitations 要用正向 scope-control 語言：

Current scope:
  The demo supports controlled audio-file or short-turn microphone input.

Validation layer:
  Real-time barge-in, overlap-heavy audio, and noisy far-field audio require
  separate latency and accuracy validation before production deployment.