Customer Intelligence Platform (NLP + LLM)

An end-to-end Customer Intelligence system built on ~180k Amazon product reviews to extract actionable business insights using NLP, topic modeling, sentiment analysis, and LLM-assisted labeling.

This project demonstrates how unstructured customer feedback can be transformed into executive-ready decision support, not just exploratory NLP analysis.

---

🚀 What This Project Does

• Ingests and cleans large-scale customer review data
• Performs review-level sentiment analysis
• Discovers latent themes using BERTopic + sentence embeddings
• Filters high-signal topics (≥300 reviews) to reduce noise
• Uses an LLM to generate human-readable, actionable topic labels
• Produces business-focused outputs, including:
  • Top customer pain points (negative impact)
  • Top customer delight themes (positive impact)
  • Sentiment-weighted prioritisation metrics
  • Executive-level analytics dashboard

---

📊 Executive Dashboard (Key Output)

This project includes an executive-ready dashboard that translates NLP outputs into clear business priorities.

What Customers Love

Where Customer Pain Has the Highest Impact

What the Dashboard Answers

• What customers love most, at scale
• Where negative feedback has the highest business impact
• Which issues should be prioritised first, based on impact, not raw volume
• How sentiment and topic prevalence change over time

📎 Interactive Power BI file:
dashboards/powerbi/customer_intelligence.pbix

---

🧠 Core Metrics

To move beyond raw sentiment counts, two custom metrics are introduced:

• Love Score = Positive sentiment × Topic frequency
• Pain Score = Negative sentiment × Topic frequency

These metrics enable impact-based prioritisation rather than anecdotal decision-making.

---

📊 Example Topics Identified

• Ineffective Nail Remover
• Fragrance Strength and Longevity
• Bottle Design and Shipping Issues
• Headband Sizing and Comfort Issues
• Quality vs. Price Balance

Each topic includes:

• Review volume
• Sentiment distribution
• Average rating
• Business-oriented interpretation

---

🧠 System Architecture

Raw Reviews
↓
Cleaning & Canonical Schema
↓
Sentiment Analysis
↓
BERTopic (Topic Discovery)
↓
Topic Aggregation
↓
LLM Topic Labeling
↓
Impact Metrics (Pain / Love Scores)
↓
Executive Dashboard & Summary

---

🛠️ Tech Stack

• Python, Pandas, NumPy
• BERTopic, SentenceTransformers
• PyTorch (GPU-accelerated embeddings)
• OpenAI API (LLM-based topic labeling)
• Power BI (executive analytics dashboard)

---

📄 Key Outputs

• topic_summary.csv — sentiment, ratings, and volume per topic
• topic_labels.csv — LLM-generated topic labels
• executive_summary.md — insight-driven business summary
• customer_intelligence.pbix — interactive dashboard

---

⚠️ Design Choices & Limitations

• Topics with low confidence are intentionally surfaced as Unlabeled to avoid overconfident assignments
• The goal is prioritisation and decision support, not causal inference
• LLM outputs are aggregated and thresholded to minimise hallucination risk

---

💼 Why This Matters

This project demonstrates the ability to:

• Build production-style NLP pipelines
• Combine unsupervised ML with LLM reasoning
• Translate raw text data into decision-ready business insights
• Design analytics for product, pricing, and CX stakeholders

---

👤 Author

Iman Badrooh
Data Scientist — NLP • ML • Customer Intelligence