Building a Calmer News Feed

I enjoy my daily dose of news, but lately my feed is drowning in rage-bait. The attention economy has trapped everyone — even publicly funded organisations like ABC often chase clicks. I rarely find the objective deep dives that I can actually gain value from.

So I’m attempting to build Seithi (Tamil for “News”) — an engine to filter out the noise.

Seithi scores every article on 3 dimensions of quality:

1. Objectivity — Is this verified fact, or just speculation?
2. Tone — Is it calm and measured, or designed to trigger you?
3. Depth — Is it a clickbait snippet, or a real analysis?

How it works:

Each article is scored using zero-shot classification — a technique where you ask a language model to rate how well a piece of text matches a hypothesis, without any task-specific training data.

Specifically, I’m running valhalla/distilbart-mnli-12-3 — a distilled version of Facebook’s BART-large, fine-tuned on the Multi-Genre NLI corpus. For each axis, the model evaluates a binary question: “This article is calm and measured” vs “This article is triggering” and returns a confidence score (0–1). No labelled data needed — just carefully phrased hypotheses.

The cold-start scores are imperfect, but the plan is to improve them through active feedback. Corrections feed into SetFit (Sentence Transformer Fine-tuning) — a sample-efficient method that fine-tunes a sentence-transformers model on just a few dozen examples to learn personal definitions of quality.

The result? A feed that gets smarter — and calmer — the more I use it. 🧘‍♂️

I barely have a UI and the NLP pipeline is still rough around the edges. Stay tuned — until then I’ll be raging over my regular news feeds.