AYZIT / Sales Assistant System: AI-Powered Sephora Recommendation & Price Intelligence
Summary
The AYZIT / Sales Assistant System is an AI-powered cosmetic product recommendation and price intelligence system developed using Sephora product data. This modular system integrates a content-based recommendation engine, employing TF-IDF vectorization and cosine similarity to identify products similar in ingredients and textual descriptions. It also features a machine learning-based price prediction engine, utilizing a hybrid XGBoost and Random Forest model, to estimate market value and classify products as overpriced, fair-priced, or underpriced. The architecture incorporates a LangChain agent with GPT-4o-mini, where the LLM interprets and summarizes outputs from the recommendation and price tools, rather than performing calculations directly. MLflow ensures experiment tracking and reproducibility, while a Streamlit interface provides an interactive user experience, allowing users to get recommendations, compare prices, and understand product value.
Key takeaway
For AI Engineers designing comprehensive product intelligence systems, integrating traditional machine learning for core calculations with an LLM agent for interpretation offers a robust architecture. You should consider a hybrid model for price prediction and content-based methods for recommendations, ensuring MLflow tracks all experiments. This approach allows your LLM, like GPT-4o-mini, to focus on user-friendly response generation, enhancing explainability and user trust by leveraging reliable tool outputs rather than direct computation.
Key insights
Combining recommendation and price intelligence with an LLM agent creates an explainable, data-driven cosmetic shopping assistant.
Principles
- LLMs excel as interpreters, not primary calculators.
- Hybrid ML models balance performance and stability.
- Explainable AI requires multi-signal interpretation.
Method
The system preprocesses Sephora data, applies TF-IDF for content-based recommendations, trains a hybrid XGBoost/Random Forest model for price prediction, integrates a LangChain agent with GPT-4o-mini for interpretation, and tracks experiments with MLflow.
In practice
- Use TF-IDF and cosine similarity for ingredient-based recommendations.
- Compare predicted vs. actual prices for product value insights.
- Employ LangChain agents to orchestrate ML tools with LLMs.
Topics
- AI Sales Assistant
- Product Recommendation Systems
- Price Intelligence
- LangChain Agents
- GPT-4o-mini
- MLOps
- Streamlit Applications
Code references
Best for: Machine Learning Engineer, AI Engineer, MLOps Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.