At a glance
- Live since: 13 November 2025
- Running time: 6.4 months
- Articles processed: 403 (avg 1.4/day, last 30 days)
- Production workflows: 5 n8n services, fully automated
- Approach: decompose news into neutral facts, then rebuild as three framings, not classify-and-label
- Built and operated by: Okavyx (Adelaide), one-person engineering
The problem with "bias labels"
Most bias-detection tools do the same thing: they read an article, score it on a left/centre/right axis, and stick a label on it. Useful at a glance, but it is secondary commentary on someone else's framing. The reader still only sees one version of the story, just with a sticker on it.
That has two issues. The label is somebody's opinion of bias, not a demonstration of it. And the underlying article still does its framing work on the reader before they have parsed the label.
What BiasFeed does instead
BiasFeed treats every news article as a transformation of an underlying set of facts. The system:
- Reads a candidate story.
- Decomposes it into a list of atomic, source-attributable facts. No commentary, no adjectives, no framing. Just what happened.
- Rebuilds the story three times from that same fact list: once with centrist framing, once with left framing, once with right framing. Same facts, three voices.
- Presents all three side by side on the article page.
The reader sees bias as a transformation, not a label. The mechanism is visible. This is an epistemically different claim than what classifier-style tools offer, and it is the thing the system is actually designed to do.
The architecture
Five n8n workflows run the production pipeline:
- The Scout - daily topic discovery. Ranks candidate stories from the source list, hands the most worthwhile one to the Researcher.
- The Researcher - the core of the system. Decomposes the chosen article into atomic facts, then generates the three framings.
- Perplexity - research handoff. Pulls source material to feed the Researcher.
- RSS Curator - ingestion across the feed list.
- Art Department - hybrid image picker. Wikimedia Commons for named entities, Pexels/Unsplash/Pixabay for concepts, with an AI "art director" choosing between them.
The whole thing runs on n8n and a small VPS. There is no separate "BiasFeed application", the workflows are the application.
Six months in: what the numbers say
- 403 articles processed end-to-end since launch. 346 text articles plus 57 videos.
- 42 articles in the last 30 days (avg 1.4/day). The system runs unattended.
- 5 production workflows, all currently active.
- ~7 to 8 third-party API calls per article, mostly GPT-4o (one fact decomposition, three framings, one art-director call), one Perplexity research call, one to two image API calls.
- Image source mix: Wikimedia Commons 144, Pexels 103, YouTube 50, Pixabay 45, Unsplash 38. The hybrid router behaves as designed: named entities go to Wikimedia, concepts go to stock providers.
The article rate is deliberately conservative. The system can run faster. It is not running faster because volume is not the point. A small number of well-decomposed articles is more useful than a firehose.
What this proves
Most "AI strategy" engagements stall in slide decks. BiasFeed is the opposite of that: a production pipeline, running unattended, on a real VPS, processing real news from real sources, every day, for six months. It is the kind of system Okavyx builds for clients when they hire us to do AI properly rather than buy AI properly.
The unique technical bet, decompose then rebuild, is the kind of decision that you cannot make from a vendor's brochure. It comes from having to actually live with how LLMs handle source material and bias. That is the value of working with people who build the systems themselves, not just recommend them.
Frequently asked
Is BiasFeed open to the public? Yes. Read it at biasfeed.com.
Who built it? Steve Ogilvie at Okavyx, in Adelaide. Five n8n workflows, one VPS, designed and operated end-to-end by one person.
What is the cost to run? Around AUD $15 per month all in: $10 for the VPS and roughly $5 for AI API calls via OpenRouter. The system is intentionally lightweight.
Could you build something like this for my business? Yes. This is the same skill set we apply to client work: workflow design, multi-step LLM orchestration, structured-data extraction, hybrid model and API pipelines. Get in touch.