ReelGuardian

ReelGuardian is the first AI-based "short-form content" misinformation checker. By looking at the claims made in the content, it will fact-check it against millions of reputable sources.

Comment

Inspiration

Short videos spread fast, but so do bad claims. Friends and family send clips every single day, with confident statements but no sources. We wanted a “pause → check → show sources” system that works in seconds, not hours.

What it does

User uploads a video, ReelGuardian transcribes it, pulls 3–6 fact-checkable claims, and returns verdicts: supported, refuted, or needs context. Each verdict includes a brief rationale, confidence score, and up to 3 citations, biased to primary sources (.gov/.edu, for example). ReelGuardian brings a visual summary (sampled frames from the content), so it also fact-checks based on what was on screen, not just what was said.

How we built it

Challenges we ran into

The problems and the solutions:

  • Cold starts & CUDA problems: fixed with a cron GPU warmer.
  • LLM JSON chaos: added a tolerant parser and a strict schema to keep the pipeline flowing.
  • Search noise: biased to primary sources, URL dedupe.
  • Timestamp alignment: mapping loose claims back to exact ASR segments for context.
  • API rate/latency spikes: model rotation and CPU fallback to avoid user-facing failures (OpenRouter).
  • Lack of visual information: passing sample frames as supplementary information for context.

Accomplishments that we're proud of

  • End-to-end upload → transcript → claims → verdicts with sources in one pass.

  • Solid Modal stack: cache-hot models, warmed GPU, identical behavior on web/worker.

  • Resilient LLM layer (multi-model + loose parsing) that doesn’t derail on imperfect outputs.

  • Clean, human-readable output with verdicts, confidence, and citations that people can trust.

What we learned

  • Infra simplicity wins: one pinned image + shared cache gets amazing speed

  • Warmers and caching matter more for UX than fancy prompts.

  • Primary-source (and simply sourcing the information) bias can improve trust (and judge confidence) immediately, which compensates for possible LLM slip-ups.

What's next for ReelGuardian

  • Browser extension: inline overlays on TikTok, YouTube Shorts, Reels.

  • Real-time mode for livestreams/spaces; incremental ASR + rolling verdicts.

  • Richer vision: reliable OCR for on-screen text, logo/channel provenance.

  • Proper dialogue isolation and voice detection (narrator vs dialogue, etc).

  • Multilingual ASR + diarization

  • Shareable reports and an API for moderators, creators, and newsrooms.

Built With

  • github
  • modal
  • nextjs
  • openrouter
  • tavily
  • whisper
Share this project:

Updates