Inspiration
The idea for Script Sentinel started with a simple observation: a screenplay can contain dozens of references to the real world—people, companies, locations, historical events, quotations, brands, and other factual details. Before a production moves forward, many of those references may need to be researched and reviewed.
We imagined a writer or production team sitting with a finished script and asking, “Which parts of this story do we need to investigate before we can safely move forward?” Doing that research manually can take hours, and it is easy to miss important details.
That is where Script Sentinel came from. We wanted to build an AI-powered research assistant that could read a screenplay, identify potentially important real-world references, research them using current web information, and organize the findings into something a human could actually review.
The goal was never to replace professional legal or production teams. It was to make the first pass of research faster, more organized, and evidence-driven.
What it does
Script Sentinel turns a screenplay PDF into a structured research report.
A user uploads a screenplay, and the system first analyzes the document to identify meaningful real-world references while preserving their surrounding context. Those references are then passed through an agentic workflow that determines how they should be researched.
Using Gemini, Google ADK, and Parallel Search, Script Sentinel searches for relevant and current information and connects each finding back to supporting sources.
The final output is designed around human review rather than pretending that an AI can make a definitive legal decision. Findings can be organized according to their level of attention, confidence, available evidence, and recommended next steps.
In simple terms:
Upload a script → find important references → research them → collect evidence → prioritize what needs human attention.
How we built it
We built Script Sentinel as an agentic workflow where each part of the system has a specific responsibility.
Gemini helps understand the screenplay and extract meaningful references from the uploaded document.
Google ADK provides the agent orchestration layer, allowing the research process to be handled as a structured workflow rather than a single AI prompt.
Parallel Search is used at runtime to perform web research and return source-backed information instead of relying entirely on the model's existing knowledge.
We built the application layer around a Python/FastAPI backend and connected the workflow to a web-based interface so that the entire process could be experienced as one application.
We also added practical safeguards such as reference deduplication, bounded searches, per-item error handling, and explicit handling for cases where the available evidence is insufficient.
One principle influenced the architecture throughout the project: finding no evidence should not automatically mean that something is safe or fictional. The system therefore treats uncertainty as a result that deserves human review, rather than hiding it.
Challenges we ran into
The biggest challenge was deciding how much of the decision-making should actually be automated.
At first, it sounds straightforward to ask an AI whether a screenplay reference is “safe.” But real-world clearance and research are rarely that simple. A reference can involve ownership, licensing, context, conflicting sources, or questions that require professional judgment.
That meant we had to design the system around evidence and uncertainty, rather than forcing every research item into a simple yes-or-no answer.
Another challenge was making the workflow reliable when a screenplay contains many references. Searching for everything without limits can quickly become inefficient, so we had to think carefully about deduplication, search boundaries, failure cases, and how to keep the output useful.
We also faced practical deployment and authentication challenges while setting up the full Google Cloud environment under the hackathon timeline. When some of that configuration could not be completed in time, we kept the limitation transparent and used a clearly identified fallback for the recorded demonstration rather than representing it as something it was not.
Accomplishments that we're proud of
We are proud that we built an end-to-end working concept rather than stopping at an AI prototype or a single prompt.
Script Sentinel can take a screenplay from document upload through reference extraction, agentic research, source gathering, and structured reporting.
We are especially proud of the way the project combines Gemini, Google ADK, and Parallel Search into one practical workflow. Parallel Search is not just mentioned in the documentation—it is integrated through its official SDK and used as part of the research process.
We are also proud of the decision to make the system transparent about uncertainty. Instead of generating confident-sounding answers when evidence is weak, Script Sentinel is designed to highlight situations that deserve additional human attention.
For us, that makes the project much more useful than simply producing another AI-generated summary.
What we learned
One of our biggest lessons was that building an effective agentic system is not about asking an AI to do everything.
It is about giving different components the right responsibilities.
The project taught us how much better the workflow becomes when document understanding, web research, reasoning, and reporting are separated into clear stages.
We also learned that citations are a product feature, not just a technical feature. When an AI produces a result that may influence a real decision, the user should be able to see where the information came from and understand when the system is uncertain.
Most importantly, we learned that the best AI tools are often the ones that help humans make better decisions rather than pretending to make every decision themselves.
What's next for script-sentinel
Script Sentinel is only the beginning.
The next version could track screenplay revisions and automatically identify what changed between drafts, maintain an evidence history as information on the web changes, and provide better collaboration features so researchers, producers, and reviewers can work together.
We would also like to expand the research layer to include specialized databases for ownership, rights, trademarks, historical records, and other sources that could make the research process even more comprehensive.
Our long-term vision is simple: make Script Sentinel a reliable research and triage layer between a finished screenplay and the professional clearance process.
The goal is not to replace the people making the final decisions.
The goal is to help them get to those decisions faster, with better evidence and fewer things falling through the cracks.
Log in or sign up for Devpost to join the conversation.