Reel

Why I built it

Streaming services make you browse by genre and decade, and what people have in their head is closer to “something slow and strange for a Sunday night”. I wanted to know whether a language model could take that sentence as it is, and whether it could do it fast enough to feel like search rather than a request.

What it is

A web app with a curated shelf of around 180 films. You type what you are in the mood for and posters stream in as you type, the shelf rearranging before you have finished the sentence.

How it works

Next.js on the front, and one calibrated yes-or-no question per film to TypeSafe’s Jev model, so every verdict is a judgement rather than a generation. Verdicts stream back in small batches; the client debounces typing, cancels stale requests and animates the grid. The demo was captured from the real page at real model timing.

What I learned

The design problem was the wait. Streaming partial results made two seconds feel alive, and cancelling stale requests mattered more than any visual polish. It changed how I scope AI features: decide what people see while the model is thinking before choosing the model.