Back

Dossiê: a multi-agent evidence dossier for misinformation

A multi-agent system that answers a suspicious news link with an evidence dossier instead of a verdict. I led the 5-person team, defined the architecture and integrated every agent into one pipeline.

Role
Team lead & architect
Context
AI Residency, LabLivre (UnB) × Instituto Eldorado · Challenge 1, Fake News & Disinformation
Period
~6-week challenge, 2026
Status
Completed
Links
Repository (GitHub)
  • Multi-agent pipeline
  • Parallel fan-out / fan-in
  • Scrum, 1-week sprints
  • Story mapping
  • AI coding agents (plan → execute)

Problem

Fact-checking tools usually return a verdict: true or false. A verdict replaces the reader’s own judgment.

Our team’s Essential Question was how AI can help people evaluate the reliability of information without replacing critical thinking.

Approach & Architecture

The user submits a suspicious news link or text. The system returns an evidence dossier, with no verdict, to support the user’s own investigation.

Dossiê pipeline The user submits a URL or text. The Ingestor normalizes it into a structured claim and document. Three agents then run in parallel: the Evidence agent gathers supporting and contradicting sources, the Text Analysis agent analyzes linguistic and rhetorical signals, and the Socratic agent generates questions that guide the user's reasoning. The Synthesizer merges their outputs into one evidence dossier with no verdict. Input URL or text Ingestor normalizes into a structured claim PARALLEL Evidence agent supporting andcontradicting sources Text Analysis agent linguistic andrhetorical signals Socratic agent questions that guidethe user's reasoning Synthesizer merges outputs Evidence dossier NO VERDICT
Fig. 1The pipeline fans out to three agents in parallel, then merges. Highlighted: the Ingestor, which I implemented.
  • Ingestor: normalizes the input (URL or raw text) into a structured claim and document.
  • Evidence agent: gathers supporting and contradicting sources.
  • Text Analysis agent: analyzes linguistic and rhetorical signals in the text.
  • Socratic agent: generates questions that guide the user’s own reasoning.
  • Synthesizer: merges the parallel outputs into one coherent dossier.

The whole system runs on a single MacBook Air M4 (24 GB unified memory, no dedicated GPU). We dropped a distributed multi-machine setup for simplicity.

My Role

I was the team lead and architect for a 5-person team.

  • I led the group: task split, sprints and integration. I orchestrated the people as well as the agents.
  • I defined the system architecture.
  • I integrated every agent into one pipeline.
  • I implemented the Ingestor agent. Each teammate owned one agent under this structure.

We ran Scrum with 1-week sprints and used story mapping for requirements. Development used AI coding agents in a structured plan → execute workflow.

Outcome

A working end-to-end pipeline that turns a suspicious link or text into an evidence dossier, built by five people in a ~6-week Challenge Based Learning cycle and running on one laptop.

Stack

  • Multi-agent pipeline with parallel fan-out and fan-in
  • Scrum with 1-week sprints, story mapping for requirements
  • AI coding agents in a plan → execute workflow