Inspiration
In an era of AI training and generating based on published media, the risk of accidental plagiarism has never been higher for screenwriters. I wanted to build a tool that lets creators verify the originality of their concepts against a database of existing intellectual property, ensuring that their ideas can move beyond rehashes of past stories.
What it does
ScriptAudit AI is a defensive tool for writers and studios. It takes a script concept and uses Actian VectorAI DB to perform a search across a selection of movies from the TMDB 5000 Movie Database to find existing movies with similar plot structures. Gemini 3 Flash then generates a similarity report highlighting specific areas of overlap. A Sphinx AI agent performs a strict audit on the generated report, checking format compliance and flagging any hallucinations in which the AI might be pulling from a specific existing film’s plot when attempting to describe the user's original idea.
How I built it
- Database: Uses Actian VectorAI DB to store and search 200 film plot premises for real-time similarity detection, passing the top 3 film matches to Gemini.
- Analysis/Insights: Uses Gemini 3 Flash (Preview), which provides deep reasoning to compare the user’s concept against the top 3 matched films.
- Auditing: Uses Sphinx CLI to call an autonomous agent that audits the final report to ensure it adheres to a strict, traceable schema.
- Interface: Uses a Streamlit dashboard that displays reasoning logs, allowing users to view and download both analysis and verification outputs. ## Challenges we ran into
- Distinguishing between tropes and actual plagiarism required significant fine-tuning of prompts for both Sphinx and Gemini.
- Running analyses required balancing API usage between Gemini 2.5 Flash-Lite and Gemini 3 Flash to conserve more powerful functionality tests for crucial moments.
- The Actian VectorAI DB results were sometimes incorrectly passed to the agentic layer, forcing the subsequent layers to operate on stale data. This was fixed by implementing more rigorous checks on the return values of the engine functions.
Accomplishments that I'm proud of
- Creating a functional system with a UI, with a successful multi-agent framework to verify scripts
- Low-latency retrieval from the database, preventing slow data from being a bottleneck
- Implement visual traceability with the mapping of similarity scores and Sphinx output logs
What I learned
- How to use VectorAI DB and Sphinx CLI, since I had no prior experience with either
- How to chain different types of models effectively, while linking correctly to the frontend
What's next for ScriptAudit AI
- Database Expansion: Ingesting the full 5,000+ movie TMDB dataset to provide a more comprehensive plagiarism check.
- Uploads: Moving from short idea script analyses to enabling full-length script uploads.
- Deployment: Deploying the application to a publicly accessible endpoint, since it is currently run locally.
Log in or sign up for Devpost to join the conversation.