ClaimFrame — Video Claim Extraction and Review Platform

Extracting timestamped factual claims from video with a self-hosted LLM, reviewed and exported through a fullstack app

What it does#

A fullstack app for uploading a video, extracting factual claims with start/end timestamps and a relevance score, reviewing and correcting them against the source footage, and exporting structured JSONL for downstream training or logging use.

Claim extraction#

A FastAPI backend accepts an uploaded file, an existing video path, or a YouTube URL and analyzes it against a dedicated extraction prompt to produce claims with begin/end timestamps and confidence scores. The first prototype ran on the Gemini API; the product backend since moved to a self-hosted, open-source Qwen-based model served with vLLM on UKP Lab GPUs.

Review interface#

A Next.js and React frontend shows the video and its extracted claims in a two-panel layout — clicking a claim jumps the video to its timestamp — and lets a reviewer confirm or correct each claim's relevance, building a feedback record per claim (text, timestamps, relevance, reviewer name, video ID) as they go.

Export#

Reviewed feedback is downloaded as a .jsonl file combining every claim, its correctness feedback, the video ID, and the reviewer's name, ready for training or logging pipelines.