finished for now · dec 2025 - jan 2026

Movie Recommender

I wanted better movie recommendations. I also wanted to turn the model into an app instead of leaving it in a notebook. Python does the learning, Go keeps it fast, and the site lets anyone try it.

The recommendations beat my baseline by 11.8%, and the app returns them in about 17 milliseconds.

mlgopythonnext.js

try it see the code

I like recommendation systems, but I did not want this project to end as another model trapped in a notebook. I wanted to search for a movie, get picks that actually made sense, and understand why each title showed up.

That meant building the whole path from raw data to a site someone could use. Features are built and checked in an offline machine learning pipeline, a ranking model is trained there, and a low-latency Go service returns ranked recommendations with lightweight explanations to a Next.js interface.

what it does

recommends movies for a MovieLens user_id based on historical ratings
searches movies by title and returns similar recommendations from a selected movie
enriches MovieLens data with TMDB metadata, including genres, popularity, release year, runtime, and poster URLs
returns ranked movie cards through a Next.js frontend backed by a Go API

Searching for movies in the recommender interface

Ranked movie recommendations with scores and reasons

how it works

The system is split into three pieces:

offline ML pipeline: Python scripts ingest MovieLens CSVs, enrich metadata from TMDB, build feature tables, create a training set, train a LightGBM ranking model, and export service data
online ranking service: a Go API handles candidate generation, feature lookup, ranking, and response formatting for /search, /rank, and movie detail endpoints
demo UI: a Next.js frontend lets someone search by movie title or enter a user id, then renders ranked results as movie cards

how a recommendation gets made

For user-based recommendations, the service accepts a request like:

{
  "user_id": 123,
  "k": 25
}

For movie-based recommendations, it can rank similar titles from a selected movie:

{
  "movie_id": 550,
  "k": 25
}

The response includes the ranked movies, score, poster URL, and simple reason strings such as genre matches or popularity signals. The current Go service uses a heuristic score over exported feature tables, while the LightGBM model is trained and saved offline. Wiring the model into online inference is the next step.

what I used

Python for data ingestion, feature engineering, evaluation, and LightGBM training
Go for the online ranking service
Next.js and Tailwind for the demo UI
MovieLens for ratings data
TMDB for movie metadata and posters

what surprised me

This project made me think beyond whether the model worked. I also had to decide how the training work, exported data, ranking endpoints, and interface should fit together so a recommendation could arrive quickly and still make sense.

It also made the tradeoff between model quality and serving complexity more concrete. The offline LightGBM pipeline gives a stronger ranking path, while the Go heuristic keeps the demo simple and fast until full model inference is wired into the service.