saga | Devpost

Inspiration

I built Saga because AI can create polished and confident writing in seconds while checking whether every claim is true still takes much longer. A statement can sound convincing even when it uses an outdated number, a weak source, or several articles that copied the same press release. I wanted to build an AI agent that earns trust by showing its work instead of asking people to believe another confident answer.

What it does

Saga is an AI fact checking agent for written content. I can give it a draft and it handles the full verification process. It finds the claims, plans the research, gathers evidence, looks for contradictions, checks sources, dates, and numbers, then decides what needs to change. It produces a corrected draft with citations and a tamper evident receipt showing what it verified.

How we built it

I built Saga in TypeScript as a group of focused agents. The Investigator searches for evidence. The Skeptic challenges that evidence and looks for conflicts. The Arbiter makes the final decision using only validated results. The Revision Agent rewrites claims that are false, outdated, or missing context. I also added citation checks, source quality scoring, source lineage, numerical verification, temporal checks, saved audit progress, and deterministic receipt generation. Live mode uses external research providers while Demo mode uses a separate fictional example that always produces the same result.

Challenges we ran into

The hardest challenge was preventing Saga from trusting the first plausible source it found. Several articles can look independent while actually repeating information from the same original source. I had to trace where the evidence came from before treating it as confirmation. I also had to make sure every correction was readable and grounded in evidence instead of producing vague warnings or placeholder text. Another important challenge was keeping Live and Demo modes completely separate so a failed Live audit could never silently show fixed Demo results.

Accomplishments that we're proud of

I am proud that Saga completes the full process from an unverified draft to a corrected and auditable result. It does more than label a claim as true or false. It explains the decision, shows the evidence, traces the original sources, corrects the writing, and creates proof that the audit was not changed. I am also proud that the Demo is clearly labeled and repeatable instead of pretending to perform live research.

What we learned

I learned that adding more agents does not automatically make an AI system trustworthy. Clear responsibilities, strict validation, independent evidence, and deterministic checks matter much more. I also learned that the final answer is only one part of a useful fact checking system. People need to understand how the answer was reached and be able to inspect the evidence themselves.

What's next for saga

Next I want to host Saga publicly, test it with more real documents, expand the evaluation set, and improve the speed of live research. I also want to keep improving the corrected writing while making sure every change remains connected to validated evidence.