genai-customer-insights-nlp

Applying Generative AI and pre-trained Transformer models to analyze customer service logs, uncovering insights in sentiment, issue patterns, and response quality to enhance customer interaction strategies.

---

📌 Project Overview

This project analyzes customer service conversation logs from Fresh Fare Meal Kits using Generative AI (GenAI) and pre-trained NLP models from the Hugging Face Transformers library. The goal is to extract actionable insights from 20 real call logs, enabling data-driven improvements to customer interaction protocols.

The project was developed for the QuickServe Meals Analytics Team as part of a broader customer experience enhancement initiative.

---

🎯 Objectives

• Apply Generative AI models (Google's FLAN-T5) to analyze text and conversation data
• Gain proficiency in utilizing pre-trained NLP models from the Transformers library
• Understand how generative AI models learn from data to generate new content
• Design and develop AI-driven solutions to solve complex problems in customer service

---

📂 Repository Structure

genai-customer-insights-nlp/
│
├── Call_Logs.csv                    # Raw dataset: 20 customer service call logs
├── gen_ai_customer_insights.ipynb   # Main Jupyter Notebook (all 5 tasks)
├── requirements.txt                 # Python dependencies
└── README.md                        # Project documentation

---

📊 Dataset

Property	Detail
File	Call_Logs.csv
Records	20 customer service conversation logs
Structure	Single Logs column (unstructured text)
Date Range	April 3 – April 29, 2024
Company	Fresh Fare Meal Kits

Each log entry contains:

• Date and Time of the call
• Agent name and Client name
• Full dialogue from greeting to resolution

---

🛠️ Tech Stack

Tool / Library	Purpose
Python 3.12	Core programming language
pandas	Data loading and manipulation
Transformers	Pre-trained NLP models (Hugging Face)
google/flan-t5-base	Conversation summarization & reason extraction
MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli	Zero-shot cancellation detection
matplotlib	Plotting and visualization
wordcloud	Word cloud generation

---

📋 Task Breakdown

Task 1: Clean and Prepare Data

Transform unstructured call logs into a structured DataFrame by extracting:

• Date, Time, Agent, Client, and Conversation columns

Task 2: Automated Conversation Summarization

Use google/flan-t5-base to generate concise summaries of each customer conversation, capturing the main points discussed.

Task 3: Detecting Cancellation Requests

Apply zero-shot classification (MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli) to flag conversations that contain cancellation requests — labelled as cancellation or other.

Task 4: Identifying Cancellation Reasons

For flagged cancellation conversations, use FLAN-T5 to extract the underlying reasons customers are choosing to cancel their subscriptions.

Task 5: Word Cloud Visualization

Generate word clouds for both full conversations and their summaries to visually identify common themes and customer concerns.

---

⚙️ Setup & Installation

Prerequisites

• Python 3.12+
• Git
• VS Code (with Jupyter extension) or GitHub Codespaces

Step 1: Clone the Repository

git clone https://github.com/your-username/genai-customer-insights-nlp.git
cd genai-customer-insights-nlp

Step 2: Create and Activate Virtual Environment

# Create virtual environment
python -m venv venv

# Activate (Linux/Mac/Codespaces)
source venv/bin/activate

Step 3: Install Dependencies

pip install --upgrade pip
pip install transformers torch sentencepiece pandas numpy matplotlib seaborn scikit-learn jupyter ipykernel notebook tqdm datasets wordcloud

Step 4: Register the venv as a Jupyter Kernel

python -m ipykernel install --user --name=venv --display-name "Python (venv)"

Step 5: Generate requirements.txt

pip freeze > requirements.txt

Step 6: Run the Notebook

Open gen_ai_customer_insights.ipynb in VS Code or Jupyter and select the Python (venv) kernel.

---

🔑 Key Findings

• 5 out of 20 conversations involved subscription cancellation requests
• Primary cancellation reasons: delivery delays, poor ingredient quality, missing recipe cards
• Most frequent issues: missing ingredients, damaged items, late deliveries
• Word cloud analysis reveals high frequency of terms: delivery, cancel, ingredient, kit, issue

---

📚 What You'll Learn

• Generative AI (GenAI) — How models like FLAN-T5 generate text from prompts
• Natural Language Processing (NLP) — Text classification, summarization, information extraction
• AI Applications — Real-world customer service analytics
• Transformers Library — Loading and using pre-trained models from Hugging Face

---

📝 Requirements

• Python Proficiency
• Basic Machine Learning Knowledge

---

🤝 Acknowledgements

• Google for the FLAN-T5 model
• MoritzLaurer for the multilingual zero-shot classification model
• Hugging Face for the Transformers library
• Fresh Fare Meal Kits / QuickServe Meals analytics team for the dataset and project brief

---

Built with ❤️ for the QuickServe Meals Analytics Team