Project / Recommender system

Tell it what you love. It finds what's next.

A sequential two-tower model trained on 15.5 million MovieLens ratings. It reads the films you pick, in order, and ranks all 56,330 titles in a few milliseconds.

Try the demo

01 / Live demo

Build a watchlist.

Every pick re-ranks the catalog, right in your browser. Tap + on a result to feed it back in.

Step 1 / Films you love

  • Nothing picked yet. Search for a film, or start from one of the sets above.

Step 2 / Your next watch

Pick a few films and recommendations appear here, ranked against the whole catalog.

02 / How it works

Three steps, one dot product.

01

Read the sequence

A causal Transformer reads the films you picked, in order, and compresses them into a single taste vector. No account or user id is needed.

02

Score every film

That vector is compared against all 56,330 films in the catalog with a dot product. Nothing is pre-filtered, so niche picks can surface.

03

Rank in milliseconds

In this demo the model runs in your browser, so there is no server to wait for. The repository also has a FastAPI service that batches concurrent requests into one forward pass.

03 / Results

Measured, not guessed.

Each user's most recent favourite is hidden, then ranked against the full catalog. 199,899 test users, no sampled shortcuts.

0.237

Recall@10

0.135

NDCG@10

5.8×

vs popularity

15.5M

Interactions

PyTorchFastAPINext.jsDockerGitHub ActionsTerraform
Read the code