Day 1 · Student handout

Domain map and interview-signal issue tree

Learners convert enterprise AI signals into a domain brief, 50 survival terms, and an issue tree for voice AI / AI Coach system delivery.

June 2026 Canonical in the 7-day tutorial Full local lesson

Day 1: 建立 domain map 與 issue tree

今日目標

你要把「我被問了很多 AI 名詞」轉成「我知道這份工作的系統邊界與交付風險」。

今日結束前,你要能說:

我正在進入的不是單一模型職位,而是 enterprise voice agent / AI Coach
系統交付。核心工作是把語音、RAG、agent、tool、policy、audit、deployment、
customer workflow 串成可以部署、驗收、維運的系統。

初學者解釋

一個企業 AI 系統可以拆成五條流:

問題例子
data flow資料從哪裡來、去哪裡、被誰保存audio、transcript、RAG chunk、tool result、audit log
permission flow誰可以看什麼、做什麼sales agent 可讀 sales SOP,但不能讀 HR 機密
responsibility flow出錯時誰負責ASR 錯名字、AI 建議違規話術、CRM 被寫錯
latency flow每一步花多久VAD 200ms、ASR 900ms、LLM TTFT 600ms、TTS 1200ms
cost flow哪一步花錢與吃 GPUASR GPU、LLM token、TTS inference、vector DB、logging storage

如果你只會講模型,面試官會擔心你只能做 demo。如果你能講這五條流,對方會知道你開始用系統工程方式思考。

End-to-end 架構圖

User / employee / salesperson
  |
  v
Audio capture
  |
  v
VAD / noise reduction / chunking
  |
  v
ASR
  |
  v
Diarization / timestamp / speaker attribution
  |
  v
Text cleaning / punctuation / hotword correction
  |
  v
Input PII gate
  |
  v
RAG retrieval / metadata filter / reranker
  |
  v
Agent orchestrator
  |-- tool registry
  |-- policy engine
  |-- memory manager
  |-- human approval queue
  |-- CRM / LMS / HRD / ERP connectors
  v
LLM output / coaching feedback / report
  |
  v
Output guardrail / citation / audit log
  |
  +--> dashboard
  +--> TTS response
  +--> enterprise write-back
  |
  v
Monitoring / evaluation / red teaming / cost tracking

50 survival terms

TermLayer影響什麼決策常見誤解
ASRvoice語音轉文字品質以為只要模型大就準
TTSvoice回覆能不能即時且自然只看聲音像不像,不量 latency
VADvoice何時開始/停止送 ASR只用分貝門檻
diarizationvoice誰在何時說話以為等同 ASR
speaker embeddingvoicespeaker clustering以為中文英文差異最大
overlap speechvoice同時講話如何歸屬以為切段就能解
hotwordvoice專有名詞是否辨對以為 post-processing 一定安全
contextual biasingvoiceASR 對 domain terms 的偏好以為 prompt 就能解所有 ASR
WER / CERvoice eval語音辨識錯誤率只看平均,不看關鍵詞
DER / JERdiarization evalspeaker 分段品質忽略 overlap
p50 / p95 latencyops平均與尾端延遲只報一次 demo 體感
embeddingRAG文件向量化以為 embedding 等於理解
vector DBRAG如何快速找相似文件以為資料放進去就完成
chunkingRAGretrieval 粒度亂切導致 citation 不準
metadataRAG權限、時效、情境過濾當成裝飾欄位
hybrid searchRAGkeyword + vector recall忽略 exact term
rerankerRAG候選文件重新排序以為 top-k 就是最終答案
top-kretrieval取幾筆候選文件跟 top-p 混淆
top-pgenerationLLM sampling拿來講 retrieval
thresholdRAG低分是否拒答沒資料也硬答
abstainRAG不足證據時拒答以為拒答是失敗
faithfulnessRAG eval回答是否忠於來源只看文字流暢
citationRAG eval來源可追溯只貼來源但內容對不上
agent orchestratoragent多步驟流程控制以為 agent 等於 chatbot
tool useagent是否呼叫外部能力忽略權限與副作用
tool registryagent governance哪些工具可用工具散落在 prompt 裡
schema validationagent governance參數是否可執行讓模型自由填
idempotencybackend重試是否造成重複動作不考慮副作用
dry-runbackend高風險操作先模擬直接寫入 CRM
human-in-the-loopgovernance何時人工核准只寫在 prompt
memory scopegovernance記憶能被誰讀跨 agent 洩漏
provenancegovernance資料來源追蹤事後無法稽核
AI Gatewaygovernance模型流量總控制點以為只是 proxy
policy enginegovernanceallow / deny / review規則散在程式碼
RBACsecurityrole-based controlrole 等於 permission
ABACsecurityattribute-based control忽略環境與資源屬性
audit loggovernance事後可重建行為只記 final answer
DLPsecurity防止資料外洩只做 output filter
PII detectionsecurity個資偵測只掃 user input
redactionsecurity個資遮罩日誌仍保存原文
prompt injectionsecurity惡意文字改變模型行為以為 system prompt 足夠
data exfiltrationsecurity資料被誘導洩漏只防外部攻擊者
OWASP LLM Top 10securityLLM 風險分類當成考試名詞
NIST AI RMFgovernanceAI 風險治理語言當成法規
red teamingsecurity eval主動攻擊系統找風險手動問幾題壞問題
eval harnesseval測試可重現demo 看起來可以就算
Dockerdeployment打包服務以為等於 production
Kubernetesdeployment編排、擴縮、更新只會背 Pod
GPU device plugindeploymentK8s 如何看見 GPU以為 GPU 自動可排程
vLLMinferenceLLM serving / KV cache / batching以為它是 governance layer
KV cacheinferencecontext/concurrency 記憶體只估 model weights
quantizationinferenceVRAM 與品質 tradeoff只看能不能塞進 GPU

今日產出

建立 domain-brief.md 草稿,包含:

1. 一句話職務地圖
2. end-to-end workflow
3. 50 survival terms
4. issue tree
5. top 10 weakness repair goals

Issue tree 範例:

能不能交付 enterprise voice AI Coach system?
|
+-- Voice quality
|   +-- ASR WER / CER
|   +-- diarization DER / JER
|   +-- overlap speech
|   +-- hotword correction
|   +-- TTS p50 / p95 latency
|
+-- Knowledge quality
|   +-- chunking
|   +-- metadata
|   +-- hybrid search
|   +-- reranker
|   +-- citation / abstain
|
+-- Agent controllability
|   +-- tool schema
|   +-- permission
|   +-- idempotency
|   +-- human approval
|   +-- audit
|
+-- Security / governance
|   +-- PII / DLP
|   +-- prompt injection
|   +-- red-team harness
|   +-- memory scope
|   +-- policy engine
|
+-- Deployment / ops
|   +-- Docker
|   +-- K8s
|   +-- GPU sizing
|   +-- observability
|   +-- rollback
|
+-- Customer delivery
    +-- onsite workflow
    +-- acceptance criteria
    +-- milestone
    +-- handoff docs