How it works
The idea
Critics and star ratings tell you what reviewers think. Tonight's Pick reads what ordinary viewers already said online, adjusts for genre, and pairs it with Qloo's taste graph so the pick is both good and yours.
Where opinions come from
Only official APIs, with our own keys, within their terms. No scraping, no automated personal accounts.
- YouTube: comments on post-release review and reaction videos (YouTube Data API).
- Bluesky: public posts that mention the title, read-only search (AT Protocol).
- TMDB reviews and Trakt comments from people who logged watching it.
- Reddit official discussion threads, only if Reddit approves our API access.
Not used: X (paid API), TikTok, Facebook and Instagram (closed), Letterboxd (no public API), IMDb and Rotten Tomatoes (no scraping).
Filtering
We drop anything posted before release (that removes trailer hype), items under 4 words, link-only posts, bots and non-English text, exact and near-duplicates, and keep only one opinion per person per title (the longest). Up to 1,000 items per title, sampled proportionally across sources. Author handles are hashed and never stored.
Classifying
Each comment is labelled once by NVIDIA Nemotron 3 Nano (on Nebius Token Factory): is it about this title, did the writer watch it, and did they like it, feel mixed, or dislike it. Questions, jokes, plot talk and hype count as no verdict. A comment counts only if it's about the title, firsthand, has a verdict, and the model is at least 60% confident.
If the model is unavailable or over its daily budget, a smaller sentiment model (Cardiff RoBERTa) is used instead, and the score says "basic classifier". The classifier's agreement with human labels is published in the project README.
The score
The headline is the share of counted opinions that liked it. Liked, mixed and disliked always add to 100. The margin is the 95% Wilson interval. With 150+ counted opinions we show the full score; 50 to 149 is an "Early read"; under 50, "Not enough talk yet" and no number.
For ranking, we use the lower bound of that interval, so 90% from 60 opinions doesn't outrank 85% from 900. Each title is also placed against every title we've scored in its genre ("Beats 82% of horror we've scored"), shown only once a genre has 15+ titles.
Median Liked % by genre (our data)
No genre has 15 scored titles yet, so there's nothing to compare.
What Qloo does
Qloo knows taste, not reception. The agent uses it to turn vibe words into tags, find taste-matched candidates, rank a shortlist against your likes, explain fit with shared tags, spot what's rising, and for two people, compare both tastes. We then rank those candidates by how real viewers felt. Every pick must come from a tool result, and the server rejects any answer that doesn't, or that includes a title you've seen.
We send Qloo only titles and tags, never names, emails or anything personal.
Limits
- People who comment online aren't a random sample of viewers.
- Titles can be review-bombed; one-opinion-per-person and post-release filtering help but don't prevent it.
- English only.
- Some big sources aren't available to us (see above), so some titles have little talk.
- Qloo affinities describe groups of people, not you personally.