Architecture: Simple Stack, Real Integrations¶

5 Docker services. 12 integration points. Zero stubs.

System diagram¶

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#1e293b', 'primaryTextColor': '#f8fafc', 'lineColor': '#64748b', 'fontSize': '13px'}} }%%
graph TB
    subgraph CLIENT ["<b>Client</b>"]
        Browser["<b>Browser</b><br/>React + Tailwind + Vite"]
    end

    subgraph GATEWAY ["<b>Gateway</b>"]
        Nginx["<b>Nginx</b><br/>:3080"]
    end

    subgraph CORE ["<b>Application Core</b>"]
        API["<b>FastAPI</b><br/>:8000"]
        LG["<b>LangGraph</b><br/>8-node Clarifier"]
    end

    subgraph DATA ["<b>Data Layer</b>"]
        PG["<b>Postgres 16</b><br/>OLTP + metrics_catalog"]
        Redis["<b>Redis 7</b><br/>cache + pub/sub"]
    end

    subgraph AI ["<b>AI & Analytics</b>"]
        Bedrock["<b>AWS Bedrock</b><br/>Sonnet 4.6 / Opus 4.7"]
        Databricks["<b>Databricks</b><br/>SQL Warehouse"]
    end

    subgraph OBS ["<b>Observability</b>"]
        Langfuse["<b>Langfuse</b><br/>LLM tracing"]
    end

    Browser -->|"HTTP + SSE + WS"| Nginx
    Nginx --> API
    API --> LG
    API --> PG
    API --> Redis
    API --> Bedrock
    API --> Databricks
    API -.->|"opt-in"| Langfuse
    Redis -->|"pub/sub"| API
    API -->|"WebSocket"| Browser

    classDef clientNode fill:#3b82f6,stroke:#1d4ed8,color:#fff,stroke-width:2px,rx:10
    classDef gatewayNode fill:#06b6d4,stroke:#0891b2,color:#fff,stroke-width:2px,rx:10
    classDef coreNode fill:#6366f1,stroke:#4f46e5,color:#fff,stroke-width:2px,rx:10
    classDef dataNode fill:#8b5cf6,stroke:#6d28d9,color:#fff,stroke-width:2px,rx:10
    classDef aiNode fill:#f59e0b,stroke:#d97706,color:#1e293b,stroke-width:2px,rx:10
    classDef obsNode fill:#64748b,stroke:#475569,color:#fff,stroke-width:1px,rx:10,stroke-dasharray:5 5
    classDef clientBox fill:#eff6ff,stroke:#3b82f6,stroke-width:2px,color:#1e293b
    classDef gatewayBox fill:#ecfeff,stroke:#06b6d4,stroke-width:2px,color:#1e293b
    classDef coreBox fill:#eef2ff,stroke:#6366f1,stroke-width:2px,color:#1e293b
    classDef dataBox fill:#f5f3ff,stroke:#8b5cf6,stroke-width:2px,color:#1e293b
    classDef aiBox fill:#fffbeb,stroke:#f59e0b,stroke-width:2px,color:#1e293b
    classDef obsBox fill:#f8fafc,stroke:#94a3b8,stroke-width:1px,color:#64748b,stroke-dasharray:5 5

    class Browser clientNode
    class Nginx gatewayNode
    class API,LG coreNode
    class PG,Redis dataNode
    class Bedrock,Databricks aiNode
    class Langfuse obsNode
    class CLIENT clientBox
    class GATEWAY gatewayBox
    class CORE coreBox
    class DATA dataBox
    class AI aiBox
    class OBS obsBox

12 real integrations¶

Every integration is implemented and tested. No stubs.

Infrastructure (1-4)Application (5-8)Frontend (9-12)

Integration	What it does	Evidence
Redis 7	KPI cache + pub/sub fan-out + widget data cache	Real-time dashboard ticks in <5s
PostgreSQL 16	OLTP + metrics catalog + dashboard layouts	Schema in `db/init.sql`
AWS Bedrock	Two-tier LLM — 4 call sites	Sonnet 4.6 (fast) + Opus 4.7 (reasoning)
Databricks	Live SQL queries — 3 Asurion metrics	Real claim data, not synthetic

Integration	What it does	Evidence
LangGraph	8-node Clarifier with HITL pause	Full state machine, streamed over SSE
SSE	Streaming Clarifier progress	Multi-turn conversation over POST
Docker Compose	5 services orchestrated	Single `make up` — zero manual setup
Langfuse	LLM observability — opt-in tracing	Every Bedrock call traced with latency

#	Integration	What it does	Evidence
	@babel/standalone	Browser-side TSX compilation	Sealed scope — React + Icons only
	sqlglot	SQL structural safety	18 tests, 25 cases — SELECT-only enforced
:material-numeric-11-circle:{ style="color: #3b82f6" }	WebSocket	Real-time dashboard KPI updates	Sub-second server push
:material-numeric-12-circle:{ style="color: #3b82f6" }	Eval Harness	LLM codegen quality gate	7 static checks + `tsc --noEmit`

One pattern, four protocols¶

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#1e293b', 'primaryTextColor': '#f8fafc', 'lineColor': '#64748b', 'fontSize': '13px'}} }%%
graph LR
    subgraph PROTOCOLS ["<b>Protocol per Purpose</b>"]
        direction TB
        REST["<b>REST</b><br/>Dashboard CRUD<br/>Widget persistence"]
        SSE["<b>SSE</b><br/>Clarifier multi-turn<br/>Streaming progress"]
        WS["<b>WebSocket</b><br/>KPI updates<br/>Sub-second push"]
        TOOL["<b>Tool-use JSON</b><br/>Bedrock calls<br/>Structured output"]
    end

    REST -->|"standard req/resp"| API2["<b>FastAPI</b>"]
    SSE -->|"streaming POST"| API2
    WS -->|"server push"| API2
    TOOL -->|"no parsing needed"| API2

    classDef restNode fill:#3b82f6,stroke:#1d4ed8,color:#fff,stroke-width:2px,rx:8
    classDef sseNode fill:#8b5cf6,stroke:#6d28d9,color:#fff,stroke-width:2px,rx:8
    classDef wsNode fill:#22c55e,stroke:#15803d,color:#fff,stroke-width:2px,rx:8
    classDef toolNode fill:#f59e0b,stroke:#d97706,color:#1e293b,stroke-width:2px,rx:8
    classDef apiNode fill:#6366f1,stroke:#4f46e5,color:#fff,stroke-width:3px,rx:10
    classDef protocolBox fill:#f8fafc,stroke:#94a3b8,stroke-width:2px,color:#1e293b

    class REST restNode
    class SSE sseNode
    class WS wsNode
    class TOOL toolNode
    class API2 apiNode
    class PROTOCOLS protocolBox

Why this matters

Each protocol is chosen for a reason — not bolted on. The same Event Context AI Action pattern works for claims, fraud, inventory, and retention. The architecture doesn't change — just the events and reason codes.

Speaker notes (45-60s)

Point to the system diagram: "Five Docker services. Twelve real integrations. No stubs."
Use the tabbed integration tables — click through Infrastructure / Application / Frontend. Don't read every row; let the count speak.
Emphasize the protocol diagram: "REST for CRUD, SSE for the Clarifier, WebSocket for real-time, Bedrock tool-use for structured LLM output. Each protocol chosen for a reason."
Land on the key line: "Same pattern works for claims, fraud, inventory, retention. The architecture doesn't change — just the events and reason codes."