Greenlight

From how the scene should feel, to where to shoot it — Greenlight reads your scene's emotion, decides real location vs VFX, then finds and confirms the perfect setup with you.

Comment

Inspiration

  • Film pre-production is historically fragmented and slow, taking weeks to bridge creative vision with logistical reality.
  • We wanted to eliminate the manual back-and-forth of scouting, permitting, and safety checks.
  • Greenlight was built to instantly translate a raw narrative script into a safe, actionable, and crew-ready production brief.

What it does

  • Intelligent Gating: Analyzes scene tone to determine if a physical real-world location is actually required, saving compute if a soundstage works.
  • Autonomous Scouting: Pulls live web intelligence to find viable locations, factoring in permit timelines and security flags.
  • Interactive Detailing: Collaborates with the filmmaker to finalize specific setups for camera, lighting, props, and crew safety.
  • Deterministic Output: Compiles a final PDF dossier featuring live weather risks and nearest emergency infrastructure.

How we built it

  • Orchestration: We built a hierarchical multi-agent network using the Google Cloud Agent Development Kit (ADK).
  • Reasoning Engine: Powered by Gemini 3.6 Flash for nuanced scene interpretation and technical detailing.
  • Real-Time Data: Integrated the Parallel Task API via Python to fetch live location intelligence and bypass LLM cutoffs.
  • Document Compilation: Used ReportLab to deterministically map structured JSON outputs into a cleanly formatted, hallucination-free PDF.

Challenges we ran into

  • Strict Schema Handoffs: Forcing the LLM to output exact JSON schemas required by the Parallel API without adding conversational filler.
  • Gating Logic: Configuring the Root Orchestrator to strictly guard downstream tools, ensuring agents only fired when absolutely necessary.
  • Formatting Instability: Parsing nested markdown delimiters out of the agent responses so the final PDF styling wouldn't break.

Accomplishments that we're proud of

  • Successfully orchestrating a 4-node architecture (1 Root Orchestrator + 3 Sub-agents) that flawlessly passes state down the pipeline.
  • Engineering intelligent safety routing—our system successfully flagged a high-risk border shoot in Kashmir and autonomously suggested a safer stand-in.
  • Delivering a complete, end-to-end working software product.

What we learned

  • Managing conversational state across multiple specialized agents requires rigid orchestration rules using Google's ADK.
  • Generative AI is perfect for creative reasoning, but traditional deterministic code is still mandatory to guarantee the structure of the final deliverable.

What's next for Greenlight

  • Cloud Deployment: Migrating from local development to Google Cloud Vertex AI Agent Engine.
  • Audio Rehearsals: Integrating the Gemini Live API to allow directors to conduct multi-modal, real-time rehearsal sessions.
  • Enterprise Tooling: Building direct export modules to industry-standard scheduling and budgeting software.

Built With

Share this project:

Updates

Submission history