Cortex Guest Service Management System

AI-Driven Conversational Guest Service Platform (Local LLM + LangChain + LangGraph)

A full-stack AI system that behaves like a real hotel/service concierge — capable of understanding orders, complaints, and follow-ups using conversational memory and structured reasoning.

---

Project Overview

This project implements an intelligent guest service management system that:

• Accepts natural language input from guests
• Understands whether the input is an order, complaint, or status update
• Reconstructs structured meaning from the entire conversation history
• Dispatches requests dynamically
• Maintains real-time system state for inspection and administration

The system uses a local LLM (Ollama – phi3) and is designed as a multi-layer architecture:

Guest Link (UI) → Admin Cortex (Dispatch) → System Internals (Memory)
↓
LangChain + LangGraph (Agent)
↓
Ollama (phi3)

---

📸 Application Screenshots

Guest Link

Admin Cortex

System Internals

---

Core Architecture

Guest Link

• Chat interface for guest interaction
• Accepts natural language input
• Displays agent responses
• Maintains session continuity via chat_id

Admin Cortex

• Operational dashboard for active service requests
• Shows:
  • Pending requests
  • Live dispatch queue
  • Resolution archive
• Allows dispatch actions

System Internals

• Memory inspection layer
• Displays:
  • All stored requests
  • Full chat history
  • Agent reasoning output (JSON)

---

🧠 AI & Reasoning Layer

Technologies Used

Component	Purpose
Ollama (phi3)	Local LLM inference
LangChain	Model invocation abstraction
LangGraph	State-based reasoning workflow
FastAPI	Backend API
React + TypeScript	Frontend UI
Vite	Frontend bundler

---

How Reasoning Works (Few-Shot Example)

Guest Input:

i need 1 plate idly and a water bottle

System Output:

{
  "category": "ORDER",
  "reconstructedRequest": "1x plate idly, 1x water bottle",
  "analysis": {
    "urgency": "LOW",
    "severity": 1
  },
  "agentResponse": "Your breakfast order has been placed."
}

• Follow-up:
  • one more idly
• System reconstructs:
  • 2x plate idly, 1x water bottle

⚠️ The system does not respond only to the last message It reconstructs meaning from the entire chat history.

---

Project Structure:

cortex/
├── backend/
│   └── ollama_service.py   # FastAPI + LangChain + LangGraph
├── frontend/
│   ├── components/
│   │   ├── GuestLink.tsx
│   │   ├── AdminCortex.tsx
│   │   └── SystemInternals.tsx
│   ├── services/
│   │   └── ollamaService.ts
│   └── App.tsx

---

Installation & Setup

1. Install Ollama

https://ollama.com/download

Pull Model

ollama pull phi3

2. Backend Setup (Python)

cd cortex/backend
python -m venv venv
venv\Scripts\activate   # Windows
pip install fastapi uvicorn langchain langgraph requests

Run backend

uvicorn ollama_service:app --reload --port 8000

Test

http://127.0.0.1:8000/docs

3. Frontend Setup (React)

cd cortex/frontend
npm install
npm run dev

Open

http://localhost:3000

---

Data Flow

Guest UI → FastAPI (/process)
             ↓
     LangGraph (state machine)
             ↓
      LangChain (LLM call)
             ↓
        Ollama (phi3)
             ↓
 Structured JSON response

---

Number of Agents & Responsibilities

Agent	Responsibility
Guest Agent	Conversational understanding
Reasoning Agent	Intent classification + reconstruction
Admin Agent	Dispatch logic
System Memory	State persistence

(implemented as logical layers, not separate LLMs)

---

Tools Used

• Ollama (local inference)
• LangChain (LLM interface)
• LangGraph (state graph)
• FastAPI (API server)
• React + TypeScript (frontend)
• TailwindCSS (UI)
• Vite (build system)

Design Choices

• Local LLM for privacy and offline execution
• State graph instead of raw prompt chaining
• No keyword rules - fully semantic reasoning
• Structured output JSON for deterministic UI
• Single API endpoint for clarity

Innovative Features

• Conversational memory
• Multi-turn request modification
• Live system memory viewer
• Agent-driven dispatch simulation
• Structured reasoning output
• LangGraph integration

Optional future upgrades:

• Vector memory
• Tool calling (inventory, DB)
• Multi-node planning agents

---

Author

Arshath Farwyz | AI Engineer

Portfolio Linkedin