Under the hood

The technical architecture

This section is currently available in English only.

The system is a pipeline of three independent services. Each does one thing well.

Three services, one result

1. FluxusSource collects articles from 1,400+ RSS feeds every four hours. It extracts full text, metadata, and images. No filtering happens here.
2. NexusMind runs every two hours. It scores each article against all five lenses using large language models. This is the heavy lifting: thousands of articles, each evaluated on 6–8 dimensions per lens. The results are published as scored datasets.
3. ovr.news builds a static site from the scored data. It summarizes articles, translates where needed, and presents the result. By this point, the hard question ("is this constructive?") is already answered.

The services don't share a database. They communicate through published data files. If one service goes down, the others keep working with cached data. No single point of failure.

What makes this hard

The dimensions sound clean on paper. In practice, the edge cases are everywhere:

• An obituary for a beloved community figure scores high on "social cohesion" but isn't constructive news. We had to learn to filter these out.
• A story about a successful clinical trial scores high on "evidence level" and "wellbeing impact," but if it's behind a paywall, we can't read enough to score it properly.
• An article about a breakthrough technology might score 9/10 on technical performance but 2/10 on benefit distribution, because only wealthy nations have access. The weighted average captures that tension.
• The same story can match multiple lenses. A community-led reforestation project is both "Recovery" and "Belonging." The system handles this gracefully: it appears in the most relevant lens, deduplicated by an editorial layer.

Summarization

Articles come from sources worldwide. We use AI to create clear summaries, so readers quickly understand what each story is about. The summarization pipeline runs Gemma 3 (27B) locally on our own hardware — no cloud dependency for the bulk of the work.

For translation, we use a tiered approach: DeepL for the highest-ranked articles, with Gemini Flash (Google) as a fallback. Article text is sent to these cloud services for translation only — no user data is involved.

For details on how we keep summaries grounded and verifiable, see why you can trust this.

Open source

The scoring filters, training pipeline, and dimension definitions are public. The site itself is built with standard open-source tools (Astro, Tailwind, SQLite).

• LLM Distillery — the scoring filter definitions and training pipeline
• Two trained filters published on Hugging Face (the rest will follow)

Images

Images on ovr.news come from four sources:

• Source images: The preview image (og:image) from the original article. We don't host these ourselves; your browser loads them directly from the source.
• Wikimedia Commons: Freely available images under the CC BY-SA 4.0 license. Attribution is always shown.
• Unsplash: Images under the Unsplash license. Attribution is shown out of respect.
• Curated library: A collection of CC-licensed images we maintain ourselves.

Last updated: April 2026