01April 2026 to presentprismrefractor.in ↗
Prism
News transparency and bias analysis
A web app that refracts any article URL, YouTube link or pasted text into a six stage analysis: sentence level claim, opinion and rhetoric classification, source provenance, live fact checking and coverage gaps.
- Period
- April 2026 to present
- Stack
- Next.js, TypeScript, React, Tailwind, Node.js, Gemini, Groq, RAG
One article, six readings
Each sentence is classed as a claim, an opinion or rhetoric. Then Prism traces where the sources come from, fact checks the claims live and shows what the coverage left out. Move the pointer to aim the beam.
What it took
93.3%28 of 30 labelled claims, zero incorrect verdicts
Measured on an offline evaluation harness built with no new dependencies. Every decisive verdict has to carry resolved citations from two independent owner groups, and every downgrade path routes to unverified instead of an adverse ruling.
18 of 18concurrent analyses held at HTTP 200
Google's 15 requests a minute ceiling is undocumented, so I worked it out from live 429 bodies. A full run now takes roughly 13 model requests, with Gemini to Groq failover, a four tier article extraction fallback and an SSRF guard on the URL fetch.
4 APIsbehind one provider agnostic call site
Five API routes orchestrate four external APIs. Swapping the model provider changes one call site, not every stage of the analysis.
Measured
Fact checks
Prism against 30 labelled claims. 28 correct, and neither miss was an incorrect verdict.
93.3%
Concurrent analyses
Prism held 18 of 18 simultaneous runs at HTTP 200 against a 15 requests a minute ceiling.
18/18