🛍️ AI Product Intelligence System

A Multimodal Product Discovery & Catalog Deduplication Engine

This repository contains the complete implementation of the AI Product Intelligence System built using pre-trained CLIP (Contrastive Language-Image Pretraining) embeddings, vector search, and custom algorithmic heuristics.

---

🎯 Project Overview & Core Tasks

📈 Task 1: Smart Product Recommendation Engine

Standard recommendation engines search for visually identical products. To improve user conversion and cross-selling, we engineered a system to recommend complementary items (e.g., matching running shoes with socks, apparel, and a fitness watch).

• Rule-Based Category Linkages: Maps search categories directly to coordinate styled accessories (e.g., Footwear $\rightarrow$ Socks, Watches, Apparel).
• Contextual Filtering: Restricts candidate recommendations by gender (e.g., Men or Unisex for male items) and usage (locking suggestions to Sports accessories when viewing running shoes).
• CLIP Visual Affinity: Ranks candidates within coordinate categories using cosine similarity of CLIP image embeddings.

📊 Task 1 Visual Results

---

🗃️ Task 2: Unique Product Catalog Creation

Marketplaces face data quality issues due to sellers uploading near-duplicate products (e.g., slightly altered names, different packages like "Pack of 2"). This pipeline automatically groups duplicates and cleanses the database into a canonical unique catalog.

• Leader-Centroid Clustering: Designates the first unvisited item as a leader and clusters all other unvisited items with cosine similarity $\ge 0.88$ (adjustable threshold).
• Medoid Selection: To select a single "canonical" product for each cluster, it computes the pairwise intra-cluster similarity matrix and identifies the medoid item (highest average similarity).
• Canonical Name Synthesis: Standardizes titles by cleaning noise suffixes, duplicate indices, pack counts (e.g. "Pack of 3", "Size L", "Combo", "Shoe A") using Regex patterns.

📊 Task 2 Visual Results

---

🔍 Task 3: Reverse Product Search

Enables natural language text search against product image databases by mapping textual semantics directly to visual assets in a shared vector space.

• Shared Space Retrieval: Maps text queries into the same joint embedding space as product images using the CLIP text encoder.
• Semantic Matching: Ranks candidate images by computing cosine similarity between the query text embedding and image embeddings.

📊 Task 3 Visual Results

---

🛠️ Installation & Setup

1. Clone the Repository:

   git clone https://github.com/khushalkks/GenAi.git
   cd GenAi

2. Install Required Libraries:

   python -m pip install streamlit sentence-transformers pandas numpy pillow scikit-learn matplotlib kagglehub

3. Run the Interactive Streamlit Web Application:

   streamlit run app.py

   This launches a browser dashboard at http://localhost:8501/.

4. Run the Self-Contained Notebook:

   • Open Bootcamp_Day2_Homework.ipynb in VS Code or Jupyter Lab.
   • Run cells sequentially to download the Kaggle dataset programmatically and execute verification.

---

👨‍💻 Author & Contributors

• Khushal - Developer & Project Lead - @khushalkks

Contributions are always welcome! Feel free to open issues or submit pull requests to make this engine even better. ⭐