Enhancing Product Search with AI: My Internship Experience

During my internship as a Machine Learning Engineer at a fast-paced e-commerce startup, I worked on building a semantic search and recommendation system to improve product discovery in a large catalog.

Objective

The primary objective of my internship project was to enhance product search and retrieval using AI. Traditional keyword-based search often produced irrelevant results, making it difficult for users to quickly find what they needed.

Our goal was to:

1. Understand user intent beyond simple keyword matching.
2. Retrieve semantically relevant products from a large catalog using embeddings.
3. Rerank results to present the most meaningful items first.

By combining machine learning, vector embeddings, and LLM-based reranking, we aimed to create a system that significantly improved user search experience.

Design

The system architecture was modular and consisted of the following components:

1. Intent Classification
   • Categorized incoming queries as:
   • Recommendation (e.g., “Suggest a fruity red wine”)
   • Informational (e.g., “What is a Pinot Noir?”)
   • Conversational (e.g., “I need a wine for dinner tonight”) - Guided the downstream retrieval and reranking logic.
2. Vector-Based Retrieval
   • Embedded all product attributes into a semantic vector space using sentence transformers.
   • Retrieved top candidate products using cosine similarity on Pinecone vector database.
3. Reranking Layer
   • Applied LLM-powered reranking (via Cohere Reranker) to refine the top results.
   • Focused on contextual relevance to the query.
4. Duplicate Filtering
   • Used Levenshtein distance and semantic checks to remove near-duplicate products.
   • Ensured cleaner, more diverse search results.
5. Deployment
   • Deployed as a serverless pipeline using AWS Lambda, integrating Pinecone and third-party APIs for reranking.
   • Designed for low-latency response in a real-time user interface

Data Processing

The data processing workflow was critical to enable semantic search:

• Product Embeddings
  • Curated titles, descriptions, and attributes for thousands of products
  • Generated dense embeddings with a pre-trained transformer
• Query Pre-processing
  • Normalized, tokenized, and embedded user queries in the same vector space
  • Enabled direct semantic similarity comparisons between queries and products.
• Filtering & Cleaning
  • Applied post-processing to remove repetitive, low-quality, or near-identical results.
  • Optimized embeddings for efficient vector search in Pinecone.

Optimization

Performance optimization was a key focus to ensure a smooth user experience:

1. Latency Reduction
   • Streamlined API calls and restructured the pipeline.
   • Reduced end-to-end response time from ~2 minutes to under 30 seconds.
2. Prompt Engineering
   • Tuned LLM prompts for contextual reranking.
   • Improved semantic relevance and reduced ambiguous results.
3. Deduplication Logic
   • Integrated string-based and semantic similarity scoring.
   • Prevented duplicate or visually identical products in top recommendations.

Evaluation

To ensure the system was effective and robust:

1. Custom Evaluation Dataset
   • Created a test set of ~250 diverse queries representing real user intent.
   • Used this for regression testing during iteration.
2. Manual Review
   • Performed visual inspection of top results for subjective relevance.
   • Collected feedback from the team to improve reranking logic.
3. Embedding Similarity Analysis
   • Leveraged CLIP-based similarity scores to evaluate alignment between product images and query intent.

Key Takeaways

This project was a complete end-to-end ML experience : from data processing and semantic retrieval to latency optimization and evaluation in production.

I learned how to:

• Build hybrid search systems combining embeddings and LLMs.
• Optimize for both accuracy and latency in production ML pipelines.
• Use vector databases and serverless architectures for scalable deployment.

This experience strengthened my skills in machine learning engineering, MLOps, and semantic search, while demonstrating how data science and domain knowledge can work together to improve user experience.