How to Transform an ML System

From a Streamlit-based recommendation system to a significantly more adaptable system utilizing a variety of MLOps tools within the stack

This article recounts my experience on a project I had worked on, and my desire to revisit it in order to change the structure, design, and to make it more functional and robust.

The idea came to me after spending a great deal of time working on more significant projects at work. I had implemented systems with a true production logic and relevant MLOps benefits.

As previously mentioned, I embarked on a personal project which involved developing a recommendation system using data from Letterboxd users as well as the TMDB API. Being a user of Letterboxd, a collaborative film review platform, I wanted to receive personalized recommendations based on my past reviews. Knowing that this feature was not available on the platform, I created a basic model using content-based filtering, which compared the currently popular movies on the platform with a representation of a user’s reviews to generate a top n list of most similar films.

The recommendation system wasn’t overly complex, I deployed it using Streamlit, which allowed me to not worry too much about the intricacies. However, it remained quite…