Mental Model

Overview

One visitor turn is easiest to hold as five layers. AI sits in Ground and Converse (and emotion labeling in Express); Capture and Remember are product and storage. Anthropic names from Building effective agents (PDF).

Applied workflows

• Routing. Typed text and mic audio both land on the same Conversational Twin /chat path; voice is transcribed first via Whisper.
• Prompt Chaining. Load session → build persona prompt → Bedrock converse → save session → emotion face → optional TTS.
• Not used here (yet). Parallelization, Orchestrator-Worker, Evaluator-Optimizer.

Layers

• 1 Capture. Visitor types in the widget or speaks; mic audio hits POST /transcribe (Whisper) before chat text is ready.
• 2 Remember. Lambda loads (or creates) session JSON from Memory S3 by session_id.
• 3 Ground. Persona files (facts, LinkedIn, style, summary) are assembled into the system prompt with history.
• 4 Converse. Bedrock Nova returns reply text (and often an emotion label in JSON).
• 5 Express. Session is saved; face label is resolved; UI shows reply + SVG; optional POST /tts speaks the reply.

Layer map

Layer stack for one message. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    person[Website visitor]
    capture["1 Capture
type or Whisper"]
    remember["2 Remember
Memory S3"]
    ground["✦ 3 Ground
persona prompt"]
    converse["✦ 4 Converse
Prompt Chaining"]
    express["5 Express
face and TTS"]

    person -->|POST chat or transcribe| capture --> remember
    remember --> ground --> converse --> express
    express --> person

    classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
    classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
    classDef actor fill:#f3eef8,stroke:#6b4f9a,stroke-width:2px,color:#0d0d0d
    class ground,converse agent
    class capture,remember,express service
    class person actor

Read layers top to bottom; drill into Conversational Twin for Bedrock and voice.