Welcome to the Multimodal Recommender System Enhancement project for Netflix! This repository houses cutting-edge work aimed at improving how Netflix recommends content to its users by leveraging multiple data types, including text, images, and videos.

This project focuses on integrating state-of-the-art machine learning models across various domains such as Natural Language Processing (NLP), Computer Vision (CV), and Reinforcement Learning (RL) to build a highly efficient and personalized recommendation system. Our goal is to deliver an even more engaging user experience by providing relevant content recommendations tailored to individual preferences.

• Raw ingestion: synthetic Netflix-style viewing events stored at data/sample/synthetic_viewing_history.csv.
• Warehouse layer: DuckDB star schema with dim_users, dim_titles, fact_views, plus feature tables for engagement and popularity.
• Model training: baseline popularity and user-based collaborative filtering running against the warehouse.
• Outputs & analytics: recommendations and metrics saved to outputs/, with SQL examples under sql/ and an analysis script in analysis/.
• Data storytelling UI: React + Chart.js dashboard in frontend/ for showcasing the recommendations visually.

• Clear ETL boundaries (extract, load, transform, feature_engineering, train_models, evaluate_models) with logging and type hints in src/netflix_recommender/data_pipeline.py.
• Warehouse-style modeling in DuckDB and reusable SQL stored in sql/engagement_queries.sql for engagement analysis.
• Baseline recommenders (popularity, user-based CF) that highlight experimentation velocity on top of the warehouse.
• Automated quality via pytest (tests/), and GitHub Actions workflow (.github/workflows/ci.yml) to run the suite.
• Front-end storytelling with a React dashboard (frontend/) to showcase recommendations to stakeholders.

1. Clone & install

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

2. Run tests

pytest -q

3. Run the end-to-end pipeline

python -m netflix_recommender.run_pipeline

4. Inspect outputs
   • outputs/recommendations.csv with top titles per user and model.
   • outputs/metrics.json with precision@k and coverage.
   • analysis/recommendation_analysis.py to print a human-readable summary.

Pipeline metrics: {
  "precision_at_k": 0.5,
  "test_users": 5
}

Sample recommendations:
  user_id title_id  rank       model
0     u1      s1     1  popularity
1     u1      s3     2  popularity
2     u1      s4     3  popularity

• Navigate to frontend/, run npm install then npm run dev to launch the dashboard at http://localhost:5173.
• The dashboard reads frontend/public/sample_recommendations.json by default; replace it with outputs/recommendations.csv converted to JSON for live data.
• Components live in frontend/src/ and use Chart.js for quick visuals (coverage, precision badge, per-user tiles).

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
PIP_NO_BUILD_ISOLATION=1 pip install -e . --no-build-isolation
PYTHONPATH=src python -m netflix_recommender.run_pipeline

bash scripts/demo.sh

The demo runs the pipeline in a temporary directory, enables tracing/observability/quality checks, and prints a summary with the generated outputs.

bash scripts/verify.sh

scripts/verify.sh installs dependencies (offline-safe), runs optional format/lint/type checks if available, executes the test suite, and runs the demo as a smoke test.

• End-to-end ETL + recommendation pipeline with output artifacts in outputs/.
• Optional observability pack: structured logging, metrics registry, and OpenTelemetry scaffolding (off by default).
• Trace recorder that outputs JSONL and Markdown reports for pipeline stages (demo enables this).
• Plugin architecture with built-in engagement and cold-start tagging plugins (opt-in).
• Safety policy and data quality checks that can be enabled via environment flags.
• Pipeline reporting (summary.json and pipeline_report.md) for recruiter-friendly summaries.