All projects

CineGraph

Can an app learn my taste in films, music and food from a few swipes, and explain every pick?

CineGraph home screen: “What are we watching?” with a free-text request box and suggested queries.

Why I built it

I wanted a taste engine that explains every recommendation and runs without a server, so that my memory stays on my own device. Films, series, music, recipes, travel, cars and style, all learned from a few swipes.

What it does

A personal taste engine that runs entirely in the browser. A weekly GitHub Actions job builds a knowledge graph, text embeddings and a collaborative-filtering model from TMDB, Wikidata and MovieLens; the app then ranks, learns online from each answer and keeps its memory on the device.

Worth a look

  • Hybrid ranking: knowledge-graph paths, embeddings and implicit ALS, folded in on the device.
  • Active learning picks the next card; a logistic ranker and a Thompson bandit adapt to each answer.
  • Installable PWA with no server: static export on GitHub Pages, memory in IndexedDB.

Under the hood

  1. Weekly data build

    A GitHub Actions job pulls TMDB, Wikidata and MovieLens into Postgres with pgvector.

  2. Three kinds of signal

    A knowledge graph, text embeddings (fastembed bge-small) and a collaborative-filtering model (implicit ALS).

  3. Static export

    JSON and int8 matrices are published as a release asset and served by GitHub Pages: no backend at all.

  4. Ranking in the browser

    The PWA ranks with features, graph paths and a collaborative-filtering fold-in, all computed on the device.

  5. Learning from each answer

    Bayesian active learning picks the next card; a logistic ranker and a Thompson bandit adapt. Memory lives in IndexedDB.