কনটেন্টে যান

Enterprise AI Automations - Complete Project Summary

Privacy-first AI agents for real-world ISP operations. No cloud. No data leaks. Pure local intelligence.

Table of Contents

  1. Project Overview
  2. Architecture Philosophy
  3. System Architecture
  4. Project Modules
  5. Getting Started
  6. ISP Classifier
  7. ISP Classifier Reasoning
  8. Qwen + RAG
  9. Gemma E4B
  10. HR Assistant
  11. SLA System
  12. Enterprise Apps
  13. LLM Demos
  14. MLOps
  15. Smart Gift AI Admin
  16. Quick Start Guide
  17. Tech Stack

Project Overview

This repository contains 11 project groups, each self-contained with its own documentation and examples. The project demonstrates how local LLMs (Large Language Models) can be used for enterprise automation without cloud dependency.

Models Used

Model Parameters Purpose
Qwen 2.5 1.5B Main classification and reasoning
Gemma 4 E4B (4-bit) Efficient inference, security analysis

Architecture Philosophy

The project is built on six core principles:

Privacy First - All data stays on-premises. No cloud API calls for sensitive data. Complete data sovereignty ensures customer information never leaves your infrastructure.

Locality Only - Run entirely on your own hardware. No internet dependency. Systems work offline when needed.

Speed Matters - Small, efficient models (1.5B - 7B parameters) deliver fast inference times. Real-time responses for customer support.

Modular Design - Each project is self-contained and easy to extend. Standalone functionality means you can pick and choose what you need.

Production Ready - Built with MLOps pipelines, monitoring, and A/B testing capabilities from the start.

Human Centric - AI assists but humans decide. All decisions are explainable with complete audit trails.

System Architecture

ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé                    Input Layer                       Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé  Email Tickets  Γöé  Chat Messages  Γöé  API Calls     Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                          Γöé
                          Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé               Context & Retrieval Layer              Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé  Text Preprocessing  Γöé  Vector DB  Γöé  Knowledge    Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                          Γöé
                          Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé                  AI Processing Layer                  Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé      LM Studio  Γöé  Local LLM  Γöé  Classification     Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                          Γöé
                          Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé                Output & Review Layer                 Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé  Generated Response  Γöé  Human Review  Γöé  Customer   Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ

Project Modules

1. Getting Started

Location: getting-started/

Purpose: First steps with LM Studio and local LLM development.

What You Need: - LM Studio installed and running - Qwen 2.5 1.5B or Gemma 4 E4B model loaded - Python 3.8+ installed

Key Scripts:

Script Description
talk_to_llm.py Basic LLM communication script

Quick Start: 

import requests

response = requests.post(
    "http://localhost:1234/v1/chat/completions",
    json={
        "model": "qwen2.5-coder-1.5b-instruct",
        "messages": [{"role": "user", "content": "Hello"}]
    }
)

Key Learnings: - How to communicate with LM Studio server - Basic prompt-response patterns - Temperature and token settings

2. ISP Classifier

Location: isp-classifier/

Purpose: Customer complaint classification system using local LLMs. Maps customer complaints to diagnostic codes for efficient troubleshooting.

Author: Rakibul Hassan, Link3 Technologies

Classification Categories: - Technical Issues: Connection problems, speed issues, equipment failures - Billing: Invoice disputes, payment processing, subscription changes - Service Outages: Planned maintenance, unplanned downtime - Account Management: Profile updates, password resets, cancellations

Diagnostic Codes:

Code Description
ISP-001 ONT/Fiber issues
ISP-002 WiFi/Router issues
ISP-006 Weather-related outages
ISP-036 Fiber cut/damage
ISP-047 Signal level issues

Key Scripts:

Script Description
app-baseline-class.py Baseline rule-based classifier
app-optimized-classifiers.py Optimized classifier
app-classifier1.py to app-classifier9.py Various classifier versions
traditional_vs_ai_workflow.py Comparison script

Usage: 

from isp_classifier import LLMClassifier

classifier = LLMClassifier()
complaint = "My ONT has a red light, internet is not working"
result = classifier.classify(complaint)

Architecture: 

Customer Complaint
       Γöé
       Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé PreprocessingΓöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
       Γöé
       Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ    ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé    Rules     Γöé or Γöé     LLM      Γöé
Γöé  (Baseline)  Γöé    Γöé  (Qwen/Gemma)Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ    ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
       Γöé                  Γöé
       ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö¼ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                Γû╝
         Diagnostic Code

3. ISP Classifier Reasoning

Location: isp-classifier-reasoning/

Purpose: Adds explanation capabilities to classification. Not only classifies complaints but explains WHY it chose a particular diagnostic code.

Why Reasoning Matters:

Without Reasoning With Reasoning
"ISP-001" "ISP-001 - ONT/fiber issue detected"
No explanation "Red light pattern matches ONT failure"
Black box Transparent decision-making

Key Scripts:

Script Description
app-reasoning1.py Basic reasoning with Qwen
app-reasoning2.py Enhanced reasoning with Gemma

Example Output: 

Complaint: "My ONT has a red light and internet stopped working"

{
  "code": "ISP-001",
  "reasoning": "The 'red light' on ONT is a classic indicator of 
               fiber disconnection or ONT hardware failure.",
  "confidence": 0.92,
  "evidence": ["red light", "ONT", "internet stopped"],
  "action": "Check fiber connection at ONT, reboot ONT"
}

Benefits: 1. Transparency - Know why a decision was made 2. Trust - Operators can verify classifications 3. Debugging - Easy to find classification errors 4. Compliance - Audit trail for regulatory requirements

4. Qwen + RAG

Location: qwen-rag/

Purpose: Combines Qwen 2.5 1.5B with Retrieval-Augmented Generation for enhanced knowledge-based responses.

What is RAG? 1. Retrieving relevant documents from a knowledge base 2. Augmenting the prompt with retrieved context 3. Generating responses with accurate, up-to-date information

Architecture: 

User Query
     Γöé
     Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé  Retriever  Γöé ΓöÇΓöÇΓöÇΓöÇΓöÇΓû╢ Vector Database
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
     Γöé
     Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé   Augment   Γöé ΓöÇΓöÇΓöÇΓöÇΓöÇΓû╢ Add context to prompt
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
     Γöé
     Γû╝
ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé    LLM      Γöé ΓöÇΓöÇΓöÇΓöÇΓöÇΓû╢ Generate response
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ

Key Scripts:

Script Description
qwen_rag_demo.py Full RAG implementation
qwen_simple_rag.py Basic RAG example
qwen_vector_storage.py Vector storage utilities

Benefits:

Benefit Description
Accuracy Responses based on actual documents
Freshness Knowledge base can be updated
Attribution Sources can be cited
Hallucination Reduced by grounding in documents

Use Cases: 1. Technical Support - Pull relevant troubleshooting guides 2. Policy Q&A - Answer based on company documentation 3. Training - Provide context-aware learning materials

5. Gemma E4B

Location: gemma-e4b/

Purpose: Showcases Google's Gemma 4 E4B (4-bit quantized) model capabilities for complex reasoning and classification.

Model Specifications:

Spec Value
Model gemma-4-e4b
Quantization 4-bit
Context 8K tokens
Speed Medium
Accuracy High

Performance Comparison:

Metric Qwen 1.5B Gemma E4B
Accuracy 32.7% 58.2%
Speed (ms) 2973 4500
Context 4K 8K
Reasoning Basic Advanced

Key Scripts:

Script Description
gemma-4-e4b-app-optimized-classifiers.py Optimized classifier
gemma-4-e4b-app-reasoning2.py Reasoning classifier
gemma-4-e4b-cybersec_analysis.py Cybersecurity analysis
gemma-4-e4b-network_monitor.py Network monitoring
gemma-4-e4b-test_llm_ISP_ticket_classifier.py ISP ticket classifier
apps-standard.py Standard LLM apps
apps-slm.py SLM (Small Language Model) apps

Best Practices: 1. Use for complex classification tasks 2. Enable reasoning for transparency 3. Batch process for efficiency 4. Monitor token usage

6. HR Assistant

Location: hr-assistant/

Purpose: AI-powered HR automation tools for leave management, employee queries, and sales funnel optimization.

Components:

HR Manager - Leave Approval

Automates leave request processing and approval workflow.

HR Assistant Chatbot

Handles employee queries about policies, benefits, and procedures.

Sales Funnel AI Closer

AI-powered sales automation for converting leads.

Key Scripts:

Script Description
HR_manager_Approve_leave.py Leave approval automation
HR_Assistant.py Employee query chatbot
Link3_Sales_Funnel_AI_Closer.py Sales funnel automation

Features:

Feature Description
Leave Processing Auto-approve or flag for review
Policy Q&A Instant answers to HR questions
Lead Scoring Prioritize high-value leads
Response Generation Personalized sales outreach
Sentiment Analysis Detect employee concerns

Architecture: 

ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé                    HR Assistant                      Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö¼ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö¼ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé HR Manager  Γöé HR Chatbot  Γöé Sales Funnel AI Closer  Γöé
Γö£ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö╝ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö╝ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöñ
Γöé Leave API   Γöé Policy DB   Γöé CRM Integration         Γöé
Γöé Calendar    Γöé Benefits    Γöé Lead Database          Γöé
Γöé Team Mgmt   Γöé Procedures  Γöé Email/Telephony        Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö┤ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓö┤ΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                    Γöé
                    Γû╝
              ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
              Γöé  LLM     Γöé
              ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ

7. SLA System

Location: sla-system/

Purpose: AI-powered approval and escalation management for customer tickets. Ensures SLA compliance through intelligent automation.

Components:

SLA LLM Assistant

Monitors and manages SLA requirements in real-time.

ERP AI Approval

Automated approval system integrated with ERP workflows.

SLA Tiers:

Tier Response Time Resolution Time Examples
Critical 1 hour 4 hours Complete outage
High 4 hours 8 hours Partial connectivity
Medium 8 hours 24 hours Performance issues
Low 24 hours 72 hours General inquiries

Features:

Feature Description
Real-time Monitoring Track SLA status continuously
Auto Escalation Automatic escalation when SLA at risk
Approval Workflow Intelligent routing of approvals
Reporting SLA compliance dashboards
Integration Works with existing ticketing systems

Architecture: 

ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ     ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ     ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
Γöé   Tickets   ΓöéΓöÇΓöÇΓöÇΓöÇΓû╢Γöé  SLA Check  ΓöéΓöÇΓöÇΓöÇΓöÇΓû╢Γöé   Action    Γöé
ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ     ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ     ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ
                         Γöé                     Γöé
                         Γû╝                     Γû╝
                  ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ       ΓöîΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÉ
                  Γöé  LLM        Γöé       Γöé  Escalation Γöé
                  Γöé  Assistant  Γöé       Γöé  Manager    Γöé
                  ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ       ΓööΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÇΓöÿ

8. Enterprise Apps

Location: enterprise-apps/

Purpose: Production-ready applications for business operations, including model management, testing frameworks, and utility scripts.

Key Scripts:

Script Description
model_use_class.py Model usage and management
test_classifier.py Classifier testing framework
test_one.py Single case testing
test_llm_ISP_ticket_classifier.py ISP ticket classifier tests

Features:

Feature Description
Multi-model Support Switch between Qwen and Gemma
Usage Tracking Monitor token consumption
Performance Metrics Track accuracy and latency
Test Framework Comprehensive testing suite
Reporting Generate detailed reports

Model Manager Usage: 

from model_use_class import ModelManager

manager = ModelManager()
manager.load_model("qwen2.5-coder-1.5b-instruct")
manager.use_model("gemma-4-e4b")
response = manager.generate("What is fiber optic troubleshooting?")

9. LLM Demos

Location: llm-demos/

Purpose: Collection of demonstration scripts showcasing different LLM capabilities and use cases.

Demo Categories:

1. Basic Demos

Simple, foundational examples for beginners.

Script Description
llm_quick_demo_base.py Quick baseline demonstration
llm_mini_demo_5cases.py 5-case mini demonstration
llm_demo_small_10case.py 10-case small demonstration

2. Hierarchical Demos

Multi-level classification and decision-making examples.

Script Description
llm_hierarchical_demo.py Hierarchical classification
llm_hierarchical_class.py Class-based hierarchy

3. Stress Testing

Performance and accuracy testing under load.

Script Description
llm_stress_test_class.py Stress testing framework
llm_stress_test_report.json Test results

Performance Metrics:

Demo Cases Avg Accuracy Avg Time
Mini (5) 5 85% 2.5s
Small (10) 10 78% 3.1s
Stress (55) 55 58% 4.5s

10. MLOps

Location: mlops/

Purpose: Production-grade machine learning operations including model registry, monitoring, A/B testing, and automatic retraining.

Components:

1. Model Registry

Centralized model versioning and management.

from mlops.registry import ModelRegistry

registry = ModelRegistry("./models")
registry.register(
    model=classifier,
    version="1.2.0",
    metrics={"accuracy": 0.92, "latency": 4500}
)

2. Monitoring

Real-time model performance tracking.

from mlops.monitor import ModelMonitor

monitor = ModelMonitor()
monitor.log_prediction(
    model_id="gemma-4-e4b",
    input=complaint,
    output=code,
    latency=4500,
    confidence=0.92
)

3. A/B Testing

Compare model performance in production.

from mlops.ab_test import ABTester

tester = ABTester()
tester.create_experiment(
    name="qwen_vs_gemma",
    model_a="qwen2.5-coder-1.5b-instruct",
    model_b="gemma-4-e4b",
    traffic_split=0.5
)

4. Automatic Retraining

Trigger retraining based on performance degradation.

Features:

Feature Description
Version Control Track all model iterations
Performance Tracking Real-time accuracy monitoring
Traffic Splitting A/B test without downtime
Auto-Retraining Trigger training on degradation
Rollback Revert to previous model version

Customer Churn Prediction Example: 

from mlops.pipeline import ChurnPipeline

pipeline = ChurnPipeline()
pipeline.prepare_data("./data/customer_history.csv")
model = pipeline.train(features=["usage", "support_calls"], target="churned")
pipeline.register_model(model, version="1.0.0")
pipeline.deploy("production", version="1.0.0")

11. Smart Gift AI Admin

Location: smart-gift/

Purpose: AI-powered administration for the Smart Gift system, enabling intelligent gift matching, customer segmentation, and promotional automation.

Features:

Feature Description
Gift Matching AI-powered product recommendations
Customer Segmentation Intelligent customer grouping
Promotional Automation Automated campaign management
Analytics Real-time performance tracking

Key Scripts:

Script Description
SmartGift_AI_Admin.py Main admin interface
slm_smartgift_admin.py SLM-based admin

Usage: 

from smartgift_admin import SmartGiftAdmin

admin = SmartGiftAdmin()
segments = admin.segment_customers(data)
recommendations = admin.match_gifts(
    customer_profile=profile,
    occasion="birthday",
    budget=5000
)

Quick Start Guide

  1. Install LM Studio from https://lmstudio.ai
  2. Download a model (Qwen 2.5 1.5B or Gemma 4 E4B)
  3. Start the local server in LM Studio (localhost:1234)
  4. Run any script from the project groups

Basic Example

cd getting-started
python talk_to_llm.py

ISP Classification Example

cd isp-classifier
python app-classifier1.py

Tech Stack

Language:        Python 3.10+
LLM Runtime:     LM Studio
Vector DB:       ChromaDB
Embeddings:      sentence-transformers
Framework:       LangChain, LlamaIndex
API Server:      FastAPI, Flask
Database:        PostgreSQL, MongoDB
Monitoring:      Grafana, Prometheus
Deployment:      Docker, Kubernetes

License

MIT License

Repository: https://github.com/raqueeb/ai_work_flow
Documentation: https://aiwithr.github.io/ai_llm/