Skip to content
All work

AI systems

Fact-Verifier

A multi-agent pipeline that breaks a claim apart, checks it against five outside sources and returns a 0 to 100 truth score with citations.

Role
Personal project
Type
AI systems
Links
Source on GitHub
evidence sources
5
truth score
0–100
no GPU needed
CPU

How it fits together

  1. ClaimA sentence from the web UI or POST /api/verify
  2. DecomposeSplits the claim into parts that can be checked
  3. query
  4. RetrieveWikipedia, WolframAlpha, NewsAPI, Google Knowledge Graph and the Fact Check API
  5. evidence
  6. Analyse and compareWeighs how well the evidence supports or contradicts each part
  7. score
  8. Score and reportVerdict, truth score, confidence, reasoning steps and cited sources

The problem

A claim is easy to state and hard to check. The aim was a system that shows its work: where each piece of evidence came from, how it was weighed and how confident the verdict is.

How I approached it

  1. 1

    Decompose

    A claim is split into parts that can each be checked against evidence.

  2. 2

    Retrieve

    Agents query Wikipedia, WolframAlpha, NewsAPI, Google Knowledge Graph and the Fact Check API. Keys are optional, so it still runs with fewer sources.

  3. 3

    Analyse and compare

    Evidence from different sources is compared for agreement and contradiction.

  4. 4

    Score

    The result is a 0 to 100 truth score, a verdict and a confidence value.

What I built

  • Transparent reports: full citations, reasoning steps and an evidence breakdown for every verdict.
  • A FastAPI service with a dark, responsive web interface, running on CPU with an optional lightweight LLM.

The result

A service that returns a verdict, a truth score, a confidence value, evidence, reasoning steps and sources for any natural-language claim.

Built with

  • Python
  • FastAPI
  • Multi-agent
  • Wikipedia API
  • WolframAlpha
  • NewsAPI
  • Google Knowledge Graph
  • pytest