An AI-powered retail recommendation system that personalizes shopping through automated analysis and reasoning.

Backend: FastAPI, Python
AI: Llama NIM, Retrieval Embedding NIM, Frequency Analysis Module
Database: ChromaDB
Cloud: AWS EC2, AWS SageMaker, Docker

• 🧠 Personalized recommendations via AI coordination
• 📈 Frequency-based “Subscribe & Save” analysis
• 🔍 Semantic similarity search using embeddings
• ☁️ Scalable deployment on AWS SageMaker

(1) Initialization : The user’s purchase history data is divided into two parts: seed data and simulation data. The seed data initializes the vector database (Vector DB) with a baseline customer profile.

(2) Simulation Feed : The simulation data is streamed month by month into the frontend, representing ongoing purchase activity in real time.

(3) Backend Input : Each month’s simulation data is sent from the frontend to the FastAPI backend for processing.

(4) Frequency Analysis : The frequency analyzer identifies which products were purchased frequently enough during the current month to qualify for potential Subscribe & Save recommendations.

(5) Llama Request : The frequency analyzer sends the list of high-frequency products to the Llama NIM hosted on a SageMaker endpoint for further reasoning.

(6.1) Retrieval Query : The Llama NIM queries the Retrieval Embedding NIM to find products that are similar or related to each item the user has purchased.

(6.2) Retrieval Return : The Retrieval Embedding NIM returns all related or semantically similar products for each of the user’s purchases back to the Llama NIM.

(7.1) Vector Query : The Llama NIM then queries the Vector DB to determine whether any of the related products returned by the Retrieval NIM already exist in the user’s historical purchase data.

(7.2) Vector Return : The Vector DB returns all matches and related products from the user’s past purchase history to the Llama NIM.

(8) Reasoning : The Llama NIM reasons which items should be added to Subscribe & Save based on two criteria:

  (a) products purchased at a high frequency this month, and

  (b) similar items the user has purchased in the past, as returned by the Vector DB.

(9) Backend Return : The results of the Llama NIM’s reasoning—the final recommendations—are sent back to the FastAPI backend.

(10) Frontend Update : The SageMaker endpoint returns these Subscribe & Save recommendations to the frontend for real-time display.

(11) DB Update : The SageMaker endpoint also sends this month’s simulation data to the Vector DB, updating it so that the system’s understanding of the user’s purchasing habits evolves continuously.

Note A. The Vector DB stores and updates semantic embeddings for each user, allowing long-term personalization by tracking evolving purchase patterns.

Note B. The Frequency Analyzer determines high-frequency purchases using thresholds (e.g., purchase count per month) before passing them to the reasoning model for context-based recommendation.

Note C. The Retrieval NIM provides semantic expansion (finding related products), while the Llama NIM performs reasoning and decision-making. Together, they form a retrieval-augmented recommendation loop that continuously improves with each simulation cycle.