Overview
Guacamole is an LLM-based prototype for comparing original videos with reposted or reproduced SNS candidates. The system uses language-model reasoning to classify potential content infringement and produce a human-readable rationale.
Problem
Short-form and social-video content is often reproduced with edits, crops, captions, or format changes. The project explored whether an LLM workflow could help compare the original content and suspected reposts beyond simple metadata or exact-match detection.
Role
- Built a Python/Gradio prototype for the comparison workflow.
- Designed prompts and output structure for relation judgment and rationale generation.
- Connected video-description signals to an LLM-based classification step.
- Iterated on presentation materials and team workflow for the hackathon setting.
Result
The project was recognized with an excellence award at the FriendliAI LLM Hackathon. It demonstrated a practical product direction for LLM-assisted content-infringement review.
Materials
The presentation material combined product framing, model workflow, and example video-comparison cases. These visuals show how the prototype was explained as an LLM-agent workflow for identifying relationships between original content and reproduced or potentially infringing videos.
