Agentic Cinema

Inspiration

Film and video productions generate a huge amount of fragmented information across scripts, scenes, characters, props, wardrobe, and previous production decisions. A small continuity mistake can survive until post-production, where fixing it may require expensive VFX work or even a reshoot.

Agentic Cinema was created to explore a simple idea: what if an AI production crew could detect these problems earlier — and also remember how similar problems were handled in the past?

What it does

Agentic Cinema is an AI-powered production intelligence system built around a specialized multi-agent workflow.

When production context is submitted, four AI agents collaborate:

• Script Agent extracts structured facts about scenes, characters, props, wardrobe, locations, and story events.
• Continuity Agent compares those facts across the story timeline and detects contradictions.
• Production Memory Agent retrieves relevant historical production cases from persistent ClickHouse memory through MCP.
• Producer Agent combines current evidence with historical precedent to recommend an actionable production decision.

Possible decisions include REVIEW, FIX_IN_POST, and RESHOOT, together with severity, confidence, priority, reasoning, and recommended next steps.

The interface streams the workflow in real time, allowing users to see each specialized agent working rather than waiting for one opaque AI response.

How we built it

The system uses Google Agent Development Kit (ADK) to orchestrate specialized agents powered by Gemini.

The backend is implemented with Python and FastAPI and exposes the agent workflow through Server-Sent Events (SSE).

A Next.js, React, and TypeScript frontend provides the production mission-control interface and visualizes agent progress in real time.

For persistent production intelligence, historical production cases are stored in ClickHouse and accessed through MCP (Model Context Protocol).

The AI backend is containerized and deployed using Google Cloud Run, with Google Cloud and Vertex AI providing the production AI infrastructure.

Challenges we ran into

One of the main challenges was moving beyond a single AI prompt and designing a workflow where each agent had a clear production responsibility.

Another challenge was making historical memory useful rather than simply storing previous results. The Production Memory Agent needed to retrieve relevant precedents and make that information available to the final Producer Agent.

We also wanted the multi-agent process to be visible to the user, so we implemented real-time SSE events to expose agent progress through the interface.

Accomplishments that we're proud of

We built an end-to-end multi-agent production workflow where specialized agents can analyze screenplay information, identify continuity problems, retrieve persistent historical context, and turn those findings into an actionable production recommendation.

We are particularly proud that production memory is not just displayed as background information — it becomes part of the reasoning context used for the final production decision.

What we learned

This project showed us that agentic systems become much more useful when agents have clearly separated responsibilities and access to persistent domain-specific context.

We also learned how Google ADK, Gemini, MCP, ClickHouse, FastAPI, SSE, and a modern web interface can work together as one production-oriented AI system.

What's next for Agentic Cinema

Next, Agentic Cinema could expand beyond prop continuity into wardrobe, character state, locations, scheduling, production logistics, and budget-aware recommendations.

We also want to expand production memory so that the system can learn from a much larger history of production incidents and decisions.

The long-term goal is to turn Agentic Cinema into an AI production intelligence layer that helps production teams identify expensive problems before they reach post-production.

Built With

• adkgemini
• clickhouse
• google
• google-cloud
• mcp
• nextjs
• python
• react