কনটেন্টে যান

Architecture

AI Work Flow for Business has a deliberately small set of layers. The whole point is that you can run it on one server and reason about the whole system.

flowchart TB
    subgraph Edge
        U[Operator / End user]
    end

    subgraph App["Application layer"]
        WEB[Web UI / Chat]
        CLI[CLI / Scripts]
        API[Internal API]
    end

    subgraph Workflow["Workflow layer (the modules)"]
        M1[ISP Classifier]
        M2[SLA System]
        M3[Qwen RAG]
        M4[HR Assistant]
    end

    subgraph Core["Core layer"]
        LLM[LM Studio client]
        RAG[ChromaDB retriever]
        SET[Settings / config]
    end

    subgraph Infra["Infrastructure"]
        GPU[Local GPU server]
        LMW[LM Studio<br/>Qwen 2.5 1.5B<br/>Gemma 3 4B]
    end

    U --> WEB
    U --> CLI
    WEB --> API
    CLI --> API
    API --> M1
    API --> M2
    API --> M3
    API --> M4
    M1 --> LLM
    M2 --> LLM
    M3 --> LLM
    M3 --> RAG
    M4 --> LLM
    M4 --> RAG
    LLM --> LMW
    RAG --> SET
    LMW --> GPU

The five layers

Layer Lives in Purpose
Edge Operator's browser or terminal The person asking the question or triggering the workflow
Application Your existing systems (intranet, Slack, ticketing) Surfaces the AI to the end user
Workflow This repo's modules The actual narrow AI task (classify, route, retrieve)
Core This repo's shared libraries LM Studio client, RAG retriever, settings
Infrastructure Your server room The GPU and LM Studio runtime

Pages in this section

  • Layers — what each layer is responsible for, and what it is not
  • Data flow — what data moves where, and what never leaves your network
  • Security — threat model, network placement, audit logging

Design principles

  1. Local by default. No cloud API calls. Ever. If a module needs a model, the model is on the same network.
  2. Structured I/O. Every module has a typed input schema and a typed output schema. The model is a function, not a conversation partner.
  3. One job per module. ISP Classifier classifies. SLA Classifier assesses risk. Qwen RAG retrieves. No module does two things.
  4. Observable. Every model call is logged with input, output, latency, and token counts. You cannot operate what you cannot see.
  5. Replaceable. The LM Studio client can be swapped for any OpenAI-compatible runtime. ChromaDB can be swapped for any vector store. No module is tightly coupled to a specific tool.

See also