Inspiration

Investigations drown in paperwork. Chatbots can summarise it, but you can't cite them, and sensitive files can't go to the cloud. We wanted no answer without the page behind it, fully offline.

What it does

Alithea turns a pile of PDFs into a 3D evidence map. Every claim links to its exact quote and page and is scored by simple rules. Switch a source off to stress-test what still holds, or throw it off the map with hand gestures.

How we built it

  • Read: PyMuPDF, keeping page numbers.
  • Extract: Llama 3.1, running locally in Ollama. Every claim must quote its page word for word.
  • Score: bank 3, official document 2, witness 1, news 1. Copies of the same leak count once.
  • Map: FastAPI, SQLite, Next.js and react-force-graph-3d.
  • Hands: MediaPipe tracks each hand in the browser. A pinch is the thumb-to-index gap divided by hand size, measured in 3D.

Challenges we ran into

  • The AI invented quotes, so we check each one against the page.
  • Local AI is slow (10–20 s per page), so we cache every answer.
  • Tilted hands looked like pinches, so we measure in 3D.
  • Fast throws lost tracking, so we judge a throw by its peak speed.

Accomplishments that we're proud of

  • Every claim has a receipt: a quote and a page.
  • It runs with Wi-Fi off, camera included.
  • Throwing a source away by hand and watching the claim drop from "supported" to "weak".

What we learned

Grounding beats prompting, and simple rules earn more trust than AI confidence scores.

What's next for Alithea

Bigger cases, shared team cases, exportable reports, and more source types (email, spreadsheets, images).

Built With

Share this project:

Updates

Submission history