The Truth AI News

You usually need four sources to understand one story. Built an AI-scored news reader. Every article gets a fairness score, so the reader sees right away how slanted it is.

• 01User surface

  The whole experience the user sees

  This frame shows the live product: the only news aggregator that verifies the source, summarises it, and scores how biased the piece is. Every component is ours · scope, design, code, deploy.

• 02Stack behind the screen

  What's powering it: Next.js, TypeScript, Prisma

  5 stack components run behind this frame · Next.js, TypeScript, Prisma drive the visible UI; the rest sit in the data layer. All studio-owned.

• 03What we shipped

  Multiple sources on one page

  Fact-checked news in one place

• 04Status

  Private deploy · under NDA.

  Per the client's request the URL stays private · the build, architecture, and lessons can be shared in a scoping call.

• 01 · BRIEF

  Bias is a measurement, not an opinion.

  Defined the three score axes (bias, opinion-share, intent), trained a scoring rubric on a labelled corpus before the pipeline was built · the LLM scores against a fixed rubric, not vibes.

• 02 · ARCHITECTURE

  Stack decisions before any code.

  Decision doc captured the data flow, Next.js, TypeScript, Prisma, PostgreSQL role split, and the failure modes we'd handle in v1 vs defer. Cross-service boundaries (where AI ends and the web app begins) were drawn here so neither side leaked into the other later.

• 02 · BUILD

  Ingest → cross-source diff → score → persist.

  Prisma + Postgres schema covers articles, scores, and source-pair diffs. OpenRouter routes to the model best-fit per article-length and language. Scores stored per article version · history visible.

• 04 · POLISH

  Performance, accessibility, and observability.

  PSI / a11y / coverage budgets enforced as launch gates. Logging + metrics wired before cut-over · the team can answer 'is it working?' from a dashboard, not a Slack thread. Threat-model checklist signed off before traffic hits the box.

• 03 · SHIP

  Reader sees the score before the article.

  Article cards lead with the bias / opinion / intent ribbon · the reader knows the slant before clicking through. Cross-source diff visible on tap.

• 01Ingest

  Multi-source crawler

  Pulls articles, deduplicates, normalises into a common schema before scoring.

• 02Scoring

  Three-axis rubric · bias / opinion / intent

  Calibrated against a labelled training corpus · drift watched per release.

• 03Diff

  Cross-source comparison

  Where two outlets disagree on the same story, the diff is surfaced inline.

• 04Reader UI

  Score visible before the article

  Article cards lead with a coloured ribbon for each axis · the reader sees the slant first.

• 
  01Frame

  Feed · multi-source cards, analysis pending

  Cards show category (World), age (1m / 3m ago), source count (1-2 sources), severity (Low / Medium) and the headline · 'Analysis pending…' under each makes it transparent when the LLM hasn't finished scoring yet.

• 
  02Frame

  Article · 4-axis Signal radar

  Article synthesis on the left + Signal radar on the right with four bars: Consensus 50, Polarization 60, Velocity 70, Sentiment 50 (raw 0). Bigger area = stronger signals. The reader gets the editorial lens at a glance.

• 
  03Frame

  Tone + agreement · what the outlets share

  Tone bars (Critical 48% / Neutral 5% / Supportive 47%) plus a 'Outlets agree' list of 6 high-signal bullets distilled from the source set · the reader gets the consensus before the disagreement.

• 
  04Frame

  Article hero · 100% consensus, 0% polarization

  When all sources agree (Consensus 100, Polarization 0, Velocity 10), the radar collapses to a vertical sliver · the visual itself tells the reader 'no outlet disagreement here'. Editorial labour replaced by a glanceable shape.