Case study 01 / 04
Animu
Find the episode you left off at. Describe a scene you half remember and get the likely episode, with no spoilers past it.

Context
Animu is a small web app for people returning to a long-running show. Pick a show, describe a scene, and it names the likely episode. It saves your place in the browser, with no account.
The problem
Long shows are easy to drop and hard to pick back up. You remember a scene but not the episode, and searching for it risks spoiling everything after it. The app had to be accurate and spoiler-safe.
Approach
- 01
Each show's episodes are read once from TVmaze and Kitsu, keeping whichever source has better synopses. Every episode is embedded and stored in Postgres with pgvector, so nothing at request time calls those APIs except show search.
- 02
A description is embedded and matched to the closest episodes. When AI is configured, a re-rank step orders the top results and adds a one-line reason and a confidence, and without it the same search still works.
- 03
Spoiler safety is built in rather than filtered later. Nothing past your match is shown, upcoming episodes appear as numbers only, and each "previously on" recap is written only from that episode's own synopsis.
Outcome
Without the AI step, the right episode is in the top 3 for about three quarters of my test scenes. Popular shows are ready instantly, and any other show takes about a minute the first time. It is live at animu.net.