Multi-Model AI E-Commerce Recommender System

🔗 Live Demo: https://hybrid-rec.vercel.app/

A production-grade recommendation engine leveraging Hybrid Reinforcement Learning and Transformer Architectures to optimize user engagement and long-term value (LTV).

---

System Architecture

This project implements a Multi-Stage Recommendation Pipeline designed to handle the full user journey, from cold-start to loyal engagement.

Component	Model Architecture	Role
1. Discovery Engine	Collaborative Filtering (VAECF)	Finds latent user preferences from implicit feedback (Clicks/Views).
2. Similarity Engine	Content-Based (BERT)	Recommends "More like this" using semantic embeddings of product metadata.
3. Sequence Engine	Transformer (SASRec-style)	Predicts next-item intent based on immediate user history.
4. Decision Engine	Hybrid RL Agent (Transformer-DQN)	The Core Innovation. Optimizes the final ranking to maximize long-term reward (Purchase vs. Click).

---

Key Performance Metrics

1. Sequential Model Performance (Transformer)

Our Transformer-based sequence model significantly outperforms standard baselines, achieving 100% Catalog Coverage through a tiered inference strategy.

Metric	Score	Business Impact
Recall@10	8.72%	1.04x lift vs. Popularity Baseline. Users see more relevant items.
Coverage@50	99.40%	The model recommends niche items, preventing "Popularity Bias."
Recall@50	13.00%	Strong candidate retrieval for the RL ranking stage.

Data Source: MLflow Evaluation Logs

2. Reinforcement Learning Convergence (Sim-to-Real)

The Hybrid RL Agent was trained using a Sim-to-Real workflow in a custom Gymnasium environment.

• Stable Convergence: The agent successfully transitioned from exploration to exploitation, with epsilon decaying to 0.05.
• Reward Optimization: Cumulative reward stabilized, indicating the agent learned to prioritize high-value actions (Purchases) over low-value ones (Views).

(Epsilon Decay vs. Cumulative Reward over 400 Episodes)

---

Deployment & MLOps

Tech Stack

• Training: TensorFlow/Keras, Cornac (for VAE), Gymnasium (RL Env).
• Tracking: MLflow & DagsHub (Experiment tracking).
• Serving: FastAPI (Asynchronous inference).
• Edge Optimization: TensorFlow Lite (TFLite) for <10ms inference latency.
• Containerization: Docker for reproducible deployment.

How to Run Locally

1. Clone the Repository

   git clone [https://github.com/okefemi12/ecommerce-recommender-system.git](https://github.com/okefemi12/ecommerce-recommender-system.git)
   cd ecommerce-recommender

2. Start the API (Docker)

   docker build -t recommender-api .
   docker run -p 8000:8000 recommender-api

3. Test the Endpoint

   curl -X POST "http://localhost:8000/recommend" \
        -H "Content-Type: application/json" \
        -d '{"history": [101, 204, 305], "user_id": "user_123"}'

---

📂 Project Structure

• src/: Production FastAPI code and inference logic.
• notebooks/: Detailed Data Science experiments (EDA, Training pipelines).
• models/: Serialized model artifacts (TFLite, Keras).
• tests/: Unit and integration tests.

---