Inspiration
We were inspired by the increasing challenge of maintaining safe, compliant communities online. With millions of posts and reviews every day, it’s nearly impossible for moderators to catch every policy violation manually. We wanted to build an AI-driven tool that automates this process and saves time for moderation teams.
What we learnt
How to design and fine-tune semantic search using vector embeddings for contextual understanding (beyond simple keyword matching). How to integrate FAISS for efficient similarity search across large datasets. How to structure rule-based policies into a scalable format for AI-driven enforcement. Building fast prototypes in Python for real-world moderation scenarios.
How we built it
We used OpenAI embeddings for vector representation of both policies and user-generated content. Implemented semantic similarity search with FAISS to match reviews against policy vectors. Created a policy evaluation pipeline that flags potential violations with confidence scores. Developed test cases for different levels of violation (direct, indirect, semantic).
Challenges we ran into
Handling false positives when content was contextually similar but not actually violating. Scaling the solution for large datasets while keeping it fast for real-time moderation. Making the system understand nuance in language (e.g., sarcasm, indirect references).
Log in or sign up for Devpost to join the conversation.