Pipeline

Live ingest, content, and token use.AI;DR 0.1.18 · aa035a3

Ranking

How stories are scored

What readers read

Opens and outbound clicks, last 28 days, first-party counts

Top stories by channel

Opens, outbound clicks and outbound per open, last 7 days. Stories under 3 opens are hidden (email 10, telegram 10, extension 10).

Chrome extension

New-tab opens a day and link clicks, last 28 days, first-party counts

Pipeline

Hourly ingest, end to end

  1. FetchHN, HuggingNews, Lobsters, RSS, newsrooms, arXiv, Vietnamese
  2. ScoreJev judgments, then the LLM rubric
  3. Mergesame story collapses to one canonical item
  4. TranslateEN → VI, journalist style; VI → EN for native VI
  5. Rankimportance × quality × freshness × engagement
  6. DigestTL;DR snapshot + per-subscriber email
  7. NotifyTelegram digest + exceptional trending posts

Schedule & ratings

Cron, gates, bars

  • Hourly — Durable Object alarm drives each run (no Worker cron)
  • Watchdog — GitHub Actions at :05 / :20 / :35 / :50, 45-min coalesce
  • Hide — relevance < 0.4 never reaches the feed
  • Trending — rank in the top 0.5% of the last 72h (at least 6) and importance ≥ 7, max 3/day (6 when importance ≥ 9), 09–23h
  • Review — suggestions & submissions judged at rating ≥ 0.6
  • Flood gate — firehose feeds (arXiv, 24h newsrooms) are keyword-filtered and newest-capped per run before the scorer sees them

Source health

Per-source outcome of the last ingest run, and anything that has gone quiet. Adding a source is a data change — see the /data Sources tab and the worker README.

AnyRouter

One gateway behind every model in this pipeline

Every LLM call in the pipeline — scoring, translation, TL;DR, and the decision judges — goes through one AnyRouter key. Per-task model overrides, automatic fallback chains, and JSON mode are configured in one place instead of four.

Explore AnyRouter