PlotGuard

Script says "car is damaged," video shows it pristine. PlotGuard catches $250K mistakes before post-production does. Nova-powered continuity analysis for film & TV.

Comment

Inspiration

Every year, continuity errors cost film studios millions in reshoots and damage audience immersion. The infamous Starbucks cup in Game of Thrones, the disappearing car damage in Commando — these mistakes slip through because script supervisors manually track hundreds of details across months of production. We built PlotGuard to automate this with AI, leveraging Nova 2 Lite's unique multimodal capabilities to analyze screenplays, photos, and video in a single pipeline.

What it does

PlotGuard is a narrative intelligence platform that analyzes screenplays, production photos, and filmed footage to catch continuity errors before they become costly reshoots. It combines Amazon Nova 2 Lite's multimodal capabilities with a custom knowledge graph engine to detect:

  • Impossible knowledge: Characters referencing information they shouldn't know yet
  • Timeline conflicts: Events that contradict established chronology
  • Visual continuity errors: Costume, prop, and set differences across shooting days
  • Video-to-script mismatches: Differences between what the screenplay says and what was actually filmed

A single continuity error that makes it to theaters can cost studios $250K+ in reshoots or damage audience trust. PlotGuard catches these errors at a fraction of the cost — about $0.15 per full screenplay analysis.

How we built it

  • Frontend: Next.js 16 on AWS Amplify with Tailwind CSS and Radix UI
  • Backend: AWS Lambda (TypeScript) orchestrated by Step Functions (5-step pipeline)
  • Database: DynamoDB for analysis storage, in-memory knowledge graph (Graphology)
  • AI: Amazon Nova 2 Lite via Bedrock for text extraction, photo comparison, and native video analysis
  • Infrastructure: AWS CDK for IaC, CloudWatch + X-Ray for monitoring
  • Cost control: AWS Budgets with alerts at 50%/80%/100%

The analysis pipeline works in 5 stages:

  1. Parse — validate and chunk the screenplay
  2. Extract — Nova 2 Lite identifies characters, locations, props, timeline events
  3. Merge — deduplicate entities and build the knowledge graph
  4. Check — Nova 2 Lite + graph algorithms detect inconsistencies
  5. Finalize — generate the continuity report with severity ratings

For video analysis, we leverage Nova 2 Lite's native video understanding — no frame extraction needed. The model receives the full MP4 alongside a screenplay reference and reports visual discrepancies shot by shot.

Challenges we ran into

  • Nova's JSON output format: Nova 2 Lite occasionally produces malformed JSON. We built robust parsing with fallback extraction and retry logic with exponential backoff.
  • Video analysis requires context: We discovered Nova cannot find continuity errors from video alone. The breakthrough was our "script supervision" workflow — sending the screenplay's expected state alongside the video so Nova can compare what should be there vs. what actually appears.
  • Knowledge graph complexity: Tracking implicit information transfer (e.g., Character A overhears Character B's phone call) required building a temporal knowledge state model that goes beyond simple entity extraction.
  • Balancing cost and accuracy: Nova 2 Lite processes video at ~1fps with ~288 tokens per frame. We optimized prompts and video lengths to keep analysis under $0.01 per video while maintaining detection accuracy.

Accomplishments that we're proud of

  • 100% detection rate across 5 real movie clips with known continuity errors (Commando, Pulp Fiction, Doctor Strange)
  • Knowledge state detection that catches "impossible knowledge" errors — something no existing screenwriting tool does
  • Sub-second photo comparison for continuity checking between shooting days
  • 50x cost reduction compared to Claude Sonnet for text analysis ($0.15 vs $1.50 per screenplay)
  • Production-ready architecture with monitoring, alerting, error handling, and a $10/month budget

What we learned

  • Nova 2 Lite's multimodal capabilities enable analysis workflows that weren't possible with text-only models — native video understanding is a game-changer for production workflows
  • Knowledge graphs are essential for narrative intelligence; simple prompt-based analysis misses temporal dependencies between scenes
  • The "script supervision" pattern (expected state + observed state → diff) is broadly applicable beyond film — it works for any domain where you need to verify reality against a specification
  • Pre-computing demo results is critical for reliable presentations; never rely on live API calls during a demo

What's next for PlotGuard

  • Series bible generation — automatically create and maintain continuity bibles for TV series across seasons
  • Nova Canvas integration — generate character portraits and scene concept art from screenplay descriptions
  • Real-time collaboration — multiple script supervisors annotating simultaneously during production
  • Final Draft / Highland integration — native plugins for industry-standard screenwriting software
  • Novel mode — expand analysis to long-form fiction (books, game narratives, interactive fiction)

Built With

  • bedrock
  • cdk
  • lambda
  • next.js
  • node.js
  • nova
  • s3
Share this project:

Updates