Epistemic News

THE PROBLEM

News aggregators (Apple News, Google News) optimize for engagement, not accuracy. The signal-to-noise is terrible, and even 'balanced' aggregators show you two opinions, not the underlying facts. I wanted a reader that would tell me which claims are confirmed across sources, which are disputed, and which are one-source speculation — with explicit blind-spot detection across the political spectrum.

THE APPROACH

• Ingest → cluster → dedup pipeline: ~100 RSS sources (wire services, US mainstream, international, independent) ingested, clustered by topic, and deduplicated via Jaccard similarity and topic diversity filtering.
• Multi-perspective synthesis: a Progressive, Conservative, and Libertarian LLM agent each analyze a cluster in parallel, then a Devil's Advocate agent challenges all three for groupthink and blind spots, then a synthesis agent produces the final balanced article.
• Post-synthesis validation: Chain-of-Verification (CoVe) re-checks claims after synthesis, a hallucination cap drops low-support statements, and a confidence scorer rates each article.
• MBFC bias + factual-rating lookup on every source, with cross-reference to fact-checking sources baked into the pipeline.
• Infrastructure: Inngest for the durable pipeline, Upstash Redis for rate limiting, Supabase for storage, web-push for breaking-story alerts.