Hands-On Production Level Machine Learning Project: YouTube Shorts Recommendation System (RecSys)

When reviewing many resumes recently, especially from candidates aiming to transition into Machine Learning Engineering (MLE) roles, one common challenge stood out: the lack of industry-level ML project experience. To help bridge that gap, I’m sharing a hands-on project that simulates building a YouTube Shorts Recommendation System. This project aims to deepen your understanding of the core architecture behind video recommendation systems, while guiding you through implementing a complete end-to-end RecSys pipeline.

This project is built around a realistic YouTube Shorts AI Recommendation System scenario. It focuses on building a recommendation engine that balances real-time performance, multi-objective optimization, and recommendation diversity. From candidate generation and ranking models to online inference, A/B testing, and MLOps, this project covers the full lifecycle of a modern recommendation system.

⸻

Tech Stack

• Pandas / NumPy: For log parsing, feature engineering, and sequential behavior modeling

• Scikit-learn / TensorFlow / PyTorch: Used for training candidate generation models, ranking models, and multi-objective optimization

• FAISS / ScaNN: Implements nearest-neighbor search for trending content and similar video recall

• GNN / Transformer: Models user-video interaction graphs and contextual sequences

• FastAPI: Deploys real-time model serving with live recommendation endpoints

• TFX / Vertex AI / Redis: Powers low-latency inference and real-time feature services

• Prometheus / Grafana: Monitors performance and visualizes recommendation service metrics

• Kubeflow Pipelines: Automates training, testing, and deployment workflows

⸻

Project Modules

• Session-Aware Data Pipeline

Simulates short video watch behavior, generates session-based user activity logs and content metadata

• Candidate Generation

Recalls similar videos based on user preferences using multiple strategies: collaborative filtering, trending content, and tag similarity

• Ranking Model

Builds a Transformer-based model with a multi-task loss function to simultaneously optimize for watch time, like rate, and diversity

• Online Serving

Uses FastAPI and Redis to implement a low-latency ranking service for real-time inference

• Evaluation + A/B Testing

Supports metrics like NDCG@K, Diversity Score, and Scroll Skip Rate; includes a simulated A/B testing module to iterate and improve models

• Monitoring + MLOps

Builds a fully automated training-evaluation-deployment loop with integrated log collection and real-time service monitoring

Figure 1: Latency distribution for online inference pipeline.

⸻

Who This Project Is For

• Candidates with foundational Python and ML skills looking to specialize in recommendation systems or personalization

• Individuals preparing for MLE / RecSys / AI Infrastructure roles at tech companies

• Those seeking to build a portfolio-ready, showcaseable project for GitHub, resumes, or interviews

⸻

Key Takeaways

• Gain hands-on understanding of real-world challenges in short-video recommendation, such as cold start, multi-objective trade-offs, and latency constraints

• Learn how to implement core modules from candidate generation to ranking and real-time serving

• Get familiar with common engineering tools used in production RecSys systems (e.g., FAISS, Redis, TFX, Kubeflow)

• Deliver a complete project you can package into a resume bullet point, GitHub repo, or portfolio showcase

⸻

Additional Notes

This project is fully runnable on local machines. It includes data simulation scripts, module templates, and is designed for solo completion. It also supports API deployment, monitoring dashboards, evaluation metric visualization, and includes pre-written resume bullet points and project report templates to help you showcase your work with minimal overhead.

If you have any questions after reading through this project, feel free to reach out! For those interested in resume reviews or mock interviews, my DMs are open.