DocuForge

Pick a topic and get a fact-checked, multilingual mini-documentary with licensed public footage, music, narration, and frame-by-frame license records.

Comment

What it does

Type a topic, or pick one from live trends (Wikipedia, Google Trends, Reddit, X, Hacker News), and DocuForge produces a narrated 60–90s documentary:

  • researches the topic, writes a script and fact-checks every line
  • picks public-domain footage, with a vision model checking each shot matches the narration
  • narrates in 8 languages, adds licensed music, and lists every source in the credits
  • for object stories, generates a 3D model (Tripo3D) that explodes into its parts

## How we built it Python + FastAPI, Fireworks AI (LLM + vision model), ffmpeg, ElevenLabs/Kokoro voices, three.js + Playwright for 3D, Bright Data for social trends, deployed on Daytona with films archived to Amazon S3.

## Challenges we ran into

  • "Public domain" tags were often wrong, so we added strict license filtering.
  • Keyword search returned unrelated images until we added vision-model scoring.
  • LLMs invent facts, so a second pass checks every line and part against the sources.

## Accomplishments that we're proud of A fact-checked, scored film in about a minute, with every frame traceable to its archive and license.

## What we learned Real footage plus a second model checking the first beats pure generation.

## What's next for DocuForge More archives (Library of Congress, Smithsonian), a "director's cut" editor, and embedded provenance credentials.

Built With

Share this project:

Updates

Submission history