Actuallyrelevant.news -- AI news curation finds the news that matter to humanity

Community
, , , ,
1

I built Actually Relevant (actuallyrelevant.news), a non-commercial news curation service. It crawls 80+ sources in 5 languages daily, evaluates relevance through a multi-stage LLM pipeline, and publishes ~2% of stories that actually matter to humanity.

  • No ads, no tracking, no paywall.
  • Run as a public good.
  • Free widgets and API if you want to build on top of it.

The Pipeline

Each story passes through four LLM stages, each with its own Zod schema.

  • Pre-screening passes over ~500 articles/day in batches. Each article gets an issue area classification, a 1-10 relevance rating, and an emotion tag. Fast , cheap screening pass to avoid running full analysis on irrelevant content.
  • Full analysis has 12 structured fields including a relevance title, topic label, 40-70 word summary, key quote from the original article with attribution, analyses for 6 different relevance factors (different for each issue category), a step-by-step rating calculation, and a 20-25 word relevance summary.
  • Editorial selection pitches the most relevant stories against each other, with the LLM discarding another 50% of the remaining candidates. Editorial criteria include “prefer systemic shifts over isolated incidents”, “prefer concrete impact over speculation”, etc. Selection also prefers uplifting stories to balance negativity bias in the news.
  • Deduplication prevents publishing multiple stories about the same event. First, pgvector finds the top N nearest stories by cosine distance within a time window. Then a gpt-5-nano call evaluates each candidate against a strict “same specific event” standard. The LLM returns a structured assessment per candidate with isDuplicate boolean and reasoning. Confirmed duplicates form clusters; only the primary story reaches the public site.

The weekly newsletter adds two more LLM stages on top:

  • Newsletter selection finds two stories for each of the four main issue categories. Selection criteria emphasize complementary stories (avoid same-event overlap), tonal mix, and an uplifting story quota.
  • Editorial intro generates a 2-3 sentence warm opening that weaves together positive developments from the selected stories.

Similar flows exist for posting selected stories on Bluesky and Mastodon.

Each step is triggered by a cron job. The whole thing costs about $1 / day. Which I find incredible.

Stack

  • LLM: LangChain + OpenAI (structured output with Zod via .withStructuredOutput())
  • Backend: Express + TypeScript, PostgreSQL + pgvector (Prisma ORM)
  • Frontend: React 18 + Vite + Tailwind
  • Search: Hybrid semantic + text search (pgvector embeddings + Reciprocal Rank Fusion)
  • Scheduling: node-cron with DB-configured jobs

Try it

Live site is available at actuallyrelevant.news. Click any story to see the full relevance analysis. Methodology is disclosed on the site. See footer links for free widgets and API. Try the Positivity Slider to filter out negative news (or positive ones if you’re in a dark mood)

Stewardship

I’m looking for an organization (journalism, civic tech, EA) to take it further. Willing to hand over code and assets. See the site for more information.

Happy to answer questions about the structured output patterns or the pipeline architecture.

2 Likes
2

Excellent project. If you haven’t already, you might want to also share the project information with the Knight Foundation and its Center for the Future of News.

3

Thanks so much for the suggestion! Will look into it.