What it is
Uses cheap judgments to cut a week of feeds down to a short list.
What it does: Scores every item against a profile, rereads the best in full, and picks a balanced list.
Why it belongs in the showcase: Real numbers on volume, speed and cost, plus an honest calibration test.
Caveats: Personal-scale experiment.
Evidence
Public MIT repo created 2026-10-18. README reports a real run with the example profile over 14 topics: 8,345 feed items read, 1,063 from the last seven days judged in 1,096 calls in about a minute for $0.04. The bookmark test on 211 of the author's bookmarks in 32 categories (2026-10-18) agreed with the existing category 105/211 (49.8%) for $0.005.
Caveats
- The whole reader profile is sent with every question
- Bookmark test used category names only; usage-based folders scored near zero
- Small personal experiments, not a hosted service
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.