ScriptDoctor AI: Real-Time Screenplay Fact Check

An agentic script reader that parses screenplays with Gemini Enterprise and uses Parallel's Web MCP to instantly detect, verify, and cite real-world factual and historical inaccuracies.

Comment

Inspiration

Film production mistakes are legendary—and extraordinarily expensive. From anachronistic smartphones appearing in period dramas to inaccurate police protocols, erroneous medical terminology, or trademarked brand names slipping past script readers, factual blunders break audience immersion and expose studios to legal risks.

Traditional script coverage relies on human readers scanning hundreds of pages line-by-line, a process that is time-consuming, expensive, and prone to oversight. We realized that modern screenplays needed a dedicated Agentic Script Doctor—an intelligent assistant that doesn't just read dialogue, but autonomously cross-references every historical event, scientific claim, brand mention, and technical detail against real-time global knowledge before production begins.

What it does

ScriptDoctor AI is an autonomous screenplay analysis engine that ingests screenplays, identifies subtle factual, temporal, and legal risks, and cross-checks them against live web data. Intelligent Entity & Claim Extraction: Automatically extracts dates, locations, character actions, technical terminology, and brand mentions from standard screenplay formatted text (.pdf, .fdx, or plain text).Real-Time Fact Verification: Queries the live web to check whether claims match ground-truth reality (e.g., Did iPhones exist during a scene set in 2004? Is a cited drug dosage biologically fatal?).IP & Legal Risk Flagging: Highlights live trademarked product mentions and flags potential fair-use or clearance issues for prop departments.Annotated Script Coverage & Citations: Generates an interactive audit report mapping line-by-line script citations to verified web sources, complete with suggestion rewrites for screenwriters.

How we built it

ScriptDoctor AI combines Google Cloud’s reasoning models with real-time web retrieval infrastructure:Core AI Reasoning Engine: Built on Gemini 3.5 Flash (via Google Agent Development Kit), which processes long-context screenplay documents to extract scene intent, structural context, and explicit factual assertions.Web Grounding Infrastructure: Integrated Parallel's Search & Extract MCP Server (and Parallel Search API). Parallel allows our agent to execute objective-driven semantic search queries and extract dense, token-optimized context excerpts from live web sources to verify claims. Agentic Coordination: Implemented an orchestration pipeline that splits the script into logical scenes, parallelizes entity verification, and generates real-time remediation suggestions.Frontend & Dashboard: Built a sleek Next.js and Tailwind CSS interface that presents an interactive split-screen view: the original script on the left, and color-coded risk flags with web citations on the right.

Challenges we ran into

Contextual Granularity: Screenplays contain creative license, figurative dialogue, and intentional fantasy elements. Teaching Gemini to distinguish between artistic exaggeration and unintentional, immersion-breaking factual errors required careful prompt calibration and agent routing.

Search Query Precision: Raw script claims (e.g., "Detective pulls out a Glock 17 in 1978") needed to be automatically transformed into high-precision, objective-driven queries for Parallel Search ("When was the Glock 17 first manufactured and issued?") to retrieve accurate historical data.

Parsing Screenplay Standards: Handling complex multi-page PDF script layouts—sluglines, dialogue blocks, parentheticals, and action lines—while preserving accurate line numbers for citations.

Accomplishments that we're proud of

Zero-Hallucination Fact Verification: Achieved deterministic, source-backed verification by strictly grounding Gemini's analysis in live excerpts returned via Parallel's Search API. Near Real-Time Processing: Optimized our agent pipeline to analyze a full 120-page screenplay feature film script in under two minutes by executing batched, concurrent Parallel search requests. Clean Producer-Grade UI: Created an intuitive report dashboard that seamlessly integrates into a Hollywood coverage workflow.

What we learned

The Power of Agent-Native Search: Generic keyword search APIs flood the model's context window with noisy HTML boilerplate. Using Parallel's objective-based retrieval gave our agent compressed, token-dense excerpts that drastically improved reasoning speed and factual accuracy. Agentic Decomposition: Complex multi-step reasoning works best when split into specialized sub-agents: one for parsing screenplay formatting, one for formulating search objectives, and one for synthesizing risk audits.

What's next for ScriptDoctor AI: Real-Time Screenplay Fact Check

Direct Screenwriting Software Integration: Develop native plugins for industry-standard writing software like Final Draft and WriterDuet to offer real-time inline "grammar-check style" factual feedback while the writer types.

Multimodal Prop & Location Verification: Extend our Parallel web extraction pipeline to analyze concept art, storyboard images, and location scout photos alongside script text.

Department-Specific Export Workflows: Allow automated generation of specialized reports tailored for individual production departments (e.g., a dedicated "Clearance Report" for Legal, a "Period Authenticity Guide" for Props & Wardrobe).

Built With

Share this project:

Updates

posted an update —

ScriptDoctor AI is officially submitted to the Agentic Cinema Hackathon!

Milestone Achieved: Submission Locked In & Live! We've officially submitted ScriptDoctor AI for Agentic Cinema: The Blockbuster Hackathon! What We Built Under the Hood

  • Cognitive Reasoning Engine: Leveraged Gemini 3.5 Flash (via Google Agent Development Kit) to parse full screenplay documents, extracting structural context and temporal claims.
  • Real-Time Fact Grounding: Integrated Parallel Search & Extract MCP to execute objective-based queries against live web data, eliminating model hallucinations.
  • Interactive Producer Dashboard: Built a sleek Next.js & Tailwind CSS split-screen interface displaying real-time anachronism flags and line-by-line web citations. Check Out the Code & Pitch
  • GitHub Repository: https://github.com/Masngo/ScriptDoctor-AI A huge thank you to the organizers and judges! Drop your thoughts, questions, or feedback in the comments below.

Log in or sign up for Devpost to join the conversation.

Submission history