---

AI-Powered Competitor Intelligence Tool

Project Overview

When I started this project, my goal was to build an AI tool that could automatically find competitors and analyze how companies interact with their customers online. Here's how I did it, step by step:

---

System Architecture & Design

First, I designed the system's architecture. I decided to use two main AI agents:

• Competitor Finder Agent: Finds competitors in a given industry.
• Customer Interaction Analysis Agent: Analyzes how companies interact with customers.

This modular design allows each agent to focus on a specific task, while also enabling future expansion with more features or data sources.

System Architecture Diagram

---

Data Collection & Processing

Next, I started with data from Kaggle—a dataset of about 3 million customer support tweets. Each tweet had details like who wrote it, when, and what it said.

Processing Steps: - Used Apache Spark to process the data. - Filtered for tweets that were customer questions ("inbound") and not replies, to find the start of each conversation. - Randomly selected 10,000 conversations for manageability.

What the data looked like

---

Conversation Thread Extraction

For each starting tweet, I built the full conversation by following the reply chain. For example, if a customer tweeted "My order is late!" and the company replied, I linked those together, and kept following the thread.

• Stored conversations in Parquet format for fast access.
• Used Gemini's embedding model to convert each conversation into a vector (for semantic search).
• Saved vectors in a special database (PgVector).

Instance of an identified conversation between a user and company (AppleSupport)

Transformed data where each row represents an entire conversation thread

---

AI Agent Prompt Engineering

With the data ready, I focused on the AI agents.

Customer Interaction Analysis Agent

• Wrote a detailed prompt specifying what to look for: handling negative feedback, public/private replies, tone, helpfulness, etc.
• Step-by-step instructions: extract info → analyze → write a report with examples.

Competitor Finder Agent

• Prompted to search for top competitors in a given industry and summarize their strengths.
• Decides when to use external search tools.
• Presents results in a clear, structured way.
• Iteratively tested and refined prompts for clarity and usefulness.

---

User Interface: Streamlit Dashboard

To make everything user-friendly, I built a dashboard using Streamlit:

• Sidebar: Manage API keys.
• Tabs:
  • Analyze customer interactions
  • Find competitors
• Users can enter their own parameters and see results instantly.

Streamlit Dashboard

---

Engineering Best Practices

Throughout the project, I followed best practices in software engineering:

• Docker: Ensures consistent app deployment
• UV: Dependency management
• .env files: Secure secrets management
• Git: Version control
• Logging: For debugging and monitoring

---

Evaluation & Learnings

Once everything was working, I evaluated the system:

• Customer Interaction Analysis Agent: Delivered accurate, valuable insights.
• Competitor Finder Agent: Worked, but results were sometimes inconsistent.
• RAG Evaluation: Used the RAGAS library, showing strong context precision and recall.

Key Learnings: - Building modular AI systems - Handling big data - Designing effective prompts

Limitations: - Only uses historical Twitter data - Relies on certain APIs - Flexible for future features (e.g., live data, richer analysis tools)

---

View Full Report Here

Contact Me

• ☎️: +44 78189 61950
• 📧: prathamskk@gmail.com
• LinkedIn: www.linkedin.com/in/prathamskk/

---

Explore My Other Projects