Bi

All work

Hamsa 2026

Fact-check triage for Nepali disaster misinformation, running on open-weight AI.

Live demoCode

Hamsa's home page, with a form to send a suspicious post and recently checked posts as news cards.
raw
After Nepal's August 2026 floods, recycled and fake footage spread faster than fact-checkers could trace it by hand.
design
Explainable signals over AI guesswork. Match recycled footage by fingerprint, cite evidence in Nepali and English, and never publish without a human.
build
Qwen3-VL reads the post, pHash, SigLIP and ORB match the footage against an archive, and gpt-oss drafts a cited verdict. 17 design decisions written up.
deploy
Live, with web, Telegram and Messenger intake. On 88 real posts every fast-tracked verdict was right (7 of 7), and a case reaches the desk in about 46 seconds.

Python, FastAPI, Next.js, Qdrant, Ollama, Docker