Script-to-Screen Guardian

An multi-agent AI system that ingests film scripts and proactively prevents production disasters by predicting equipment, crew, and location risks before they happen. Lights. Camera. Protect.

Comment

🎬 Inspiration

The film industry loses billions annually to production delays. A single day of shooting can cost $250,000-$500,000 for a major production, yet most production management is still done through spreadsheets, emails, and phone calls. When a key piece of equipment breaks or weather turns bad, teams react to crises hours after they begin—often too late to prevent costly delays.

We were inspired by the chaos of film production. Watching documentaries and behind-the-scenes footage, we saw how much time and money was wasted on logistics that could have been automated. Production assistants spend hours calling rental houses, checking weather reports, and managing schedules. Meanwhile, directors and producers are forced to make creative compromises because of operational constraints.

What if an AI crew could predict these disruptions before they happen?

That's how Script-to-Screen Guardian was born—a multi-agent AI system that acts as an intelligent production assistant, protecting creative vision from logistical chaos. We wanted to bring the power of modern AI to an industry that desperately needs it, helping filmmakers focus on what they do best: telling stories.

⚙️ What It Does

Script-to-Screen Guardian is a team of specialized AI agents that work together like a real production crew:

Agent Role Technology
Script Supervisor Analyzes scripts to extract scenes, locations, props, characters, and requirements Gemini Enterprise
Location Scout Cross-references locations with weather forecasts, permits, and accessibility data Custom API integrations
Logistics Agent Queries equipment inventory, maintenance logs, and crew schedules in real-time ClickHouse Cloud
Studio Head Synthesizes risks, calculates financial impact, and recommends actions Gemini + Grafana

The Workflow:

  1. Upload a film script (PDF or text)
  2. Script Supervisor Agent extracts scene metadata using Gemini
  3. Location Scout checks weather and permits for each location
  4. Logistics Agent queries ClickHouse for equipment and crew availability
  5. Studio Head generates risk assessments and actionable recommendations
  6. Dashboard visualizes everything with traffic-light risk indicators (🔴🟡🟢)

Example Output:

SCENE 12 - EXT. ABANDONED WAREHOUSE - DAY

Risks Identified:
🔴 HIGH: Rain forecast on shoot date (Sep 10)
🟡 MEDIUM: Lighting Kit Pro booked until Sep 14
🟢 LOW: Crew available

Recommendations:
1. Reschedule from Sep 10 to Sep 15 (avoid rain)
2. Secure backup lighting kit from Studio B
3. Cost impact: $12,500 additional

🛠️ How We Built It

Technology Stack

Layer Technology
Agent Runtime Google Cloud Agent Builder + Gemini Enterprise
LLM Core Gemini Enterprise (via GEAR program)
Analytical Database ClickHouse Cloud
Visualization Grafana Cloud + Streamlit
Backend Python + FastAPI
Deployment Google Cloud Run

The Multi-Agent Architecture

We built a deterministic multi-agent pipeline using Google Cloud Agent Builder. Instead of letting LLMs decide which agent to call next (which can lead to skipped steps), we implemented a SequentialAgent pattern where each agent executes in order:

Script Upload → Script Supervisor → Location Scout → Logistics Agent → Studio Head → Dashboard

ClickHouse: The Engine That Makes It Possible

We chose ClickHouse as our analytical database because:

  • Sub-millisecond queries on equipment inventory and schedules
  • Real-time analytics with AggregatingMergeTree and materialized views
  • Scalability—the system handles 240M+ rows of data
  • MCP integration—we used the official mcp-clickhouse server for agent-database interaction

Sample ClickHouse Table Structure:

CREATE TABLE equipment_inventory (
    equipment_id UUID DEFAULT generateUUIDv4(),
    equipment_name String,
    type String,
    model String,
    status String,
    location String,
    daily_rate Decimal(10,2),
    created_at DateTime DEFAULT now()
) ENGINE = MergeTree()
ORDER BY (equipment_id, created_at);

Gemini Integration

Gemini Enterprise powers the intelligence behind our agents:

  • Script analysis: Extracting scene metadata from unstructured text
  • Context understanding: Identifying creative requirements from script descriptions
  • Recommendation generation: Producing actionable, human-readable risk assessments

Dashboard & Visualization

We built a real-time production dashboard using:

  • Grafana for professional monitoring dashboards
  • Streamlit for rapid prototyping and demo UI

Deployment

We deployed the entire system on:

  • Google Cloud Run (serverless, scales to zero, $0 cost)
  • ClickHouse Cloud (free tier with $300 credits)
  • GitHub for open-source code

🚧 Challenges We Ran Into

1. ClickHouse Authentication

Challenge: Initially struggled with ClickHouse Cloud authentication. The connection kept failing with "Authentication failed" errors.

Solution: Used clickhouse-connect with proper secure connection settings. Tested with curl first:

curl.exe --user "default:password" --data-binary "SELECT 1" \
  "https://your-host.clickhouse.cloud:8443"

Once the curl test succeeded, we integrated it into our Python client.

2. Gemini API Quota Limits

Challenge: Hit free tier limits during development—the API stopped responding.

Solution: Switched to gemini-2.5-flash-lite model, which has more generous rate limits. Also learned about token management and optimized our prompts to be more efficient.

3. Multi-Agent Orchestration

Challenge: LLM-driven agent handoffs could skip critical remediation steps. The agents would sometimes "assume" tasks were completed without actually running them.

Solution: Implemented deterministic SequentialAgent pipeline—every step executes every time, regardless of what the LLM thinks. This made the system reliable and predictable.

4. Real-Time Dashboard Data

Challenge: Grafana needed optimized queries for sub-second responses on large datasets.

Solution: Used ClickHouse projections and materialized views to pre-aggregate data. This reduced query time from seconds to milliseconds.

5. Time Management

Challenge: Building a complete multi-agent system in just a few days.

Solution: Focused on the core workflow first (Script → Analysis → Dashboard), then added features iteratively. Used quick prototyping with Streamlit for the UI.

🏆 Accomplishments We're Proud Of

Technical Achievements

  • ✅ Working multi-agent system with 4 specialized agents in a deterministic pipeline
  • ✅ ClickHouse integration with sub-150ms query latency on 240M+ rows
  • ✅ Gemini-powered script analysis that extracts structured data from unstructured text
  • ✅ Real-time dashboard with risk visualization and financial impact estimates

Industry Impact

  • ✅ 48+ hours advance notice of potential production disruptions
  • ✅ Estimated $12,500+ savings per prevented delay
  • ✅ Proactive vs. reactive production management
  • ✅ Scalable solution from independent films to major studio productions

Innovation

  • ✅ Creative use of multi-agent AI for production logistics (not just content generation)
  • ✅ Deterministic agent pipeline ensures reliability
  • ✅ Combining Gemini + ClickHouse for intelligent, fast analytics
  • ✅ Production-ready architecture that could be deployed today

Demo Success

  • ✅ Complete workflow demonstrated in 3-minute video
  • ✅ Live dashboard with real data from ClickHouse
  • ✅ Working API endpoints for equipment and crew queries
  • ✅ Professional-looking UI that judges could immediately understand

📚 What We Learned

Technical Takeaways

  • ClickHouse is incredibly fast for analytical queries—240M+ rows queried in ~114ms
  • Deterministic agent pipelines are more reliable than LLM-driven handoffs
  • Gemini's reasoning capabilities are excellent for unstructured document understanding
  • Environment variables are critical for credential management in hackathon projects
  • Test infrastructure early—pre-create accounts and API keys before coding

Hackathon Lessons

  • Start with the core workflow first, then add features iteratively
  • Document everything—makes submission much easier
  • Build for the demo, not just the product
  • Good README is as important as good code
  • Clear, simple UI beats complex, confusing UI

Industry Insights

  • Agentic AI isn't just about chatbots—it's about solving real operational problems
  • The future of filmmaking will involve AI crews, not just AI writing tools
  • Predictive analytics in production can save millions
  • LLMs + Analytical Databases is a powerful combination for enterprise applications

Collaboration Lessons

  • Clear division of roles helps move faster
  • Regular check-ins keep everyone aligned
  • Version control is essential for team projects
  • Celebrate small wins along the way

🔮 What's Next for Script-to-Screen Guardian

Immediate Improvements

  • Real-time MCP server integration for deeper ClickHouse interaction
  • Weather API integration for dynamic schedule optimization
  • Vector search using Vertex AI embeddings for failure pattern matching
  • Automated remediation—from "propose fix" to "execute fix" with human verification

Future Features

  • Multiple script support (TV series, commercials, music videos)
  • Budget impact simulation before making changes
  • Integration with production software (Movie Magic Scheduling, StudioBinder)
  • Mobile app for on-set crew access
  • Multi-language script support

Commercial Potential

Market Application Value
Independent Filmmakers Affordable production management Low-budget solutions
Major Studios Save millions on insurance and delays Enterprise licensing
Production Insurance Risk assessment tool Premium reduction
Streaming Services Optimize content delivery Release scheduling

Open Source Vision

We plan to open-source the project to help the filmmaking community build better, more efficient production workflows. This could lead to:

  • Community contributions from developers worldwide
  • Industry adoption of AI-powered production tools
  • Education for filmmakers on AI capabilities

Long-Term Vision

Imagine a world where:

  • Every script is automatically analyzed for production feasibility
  • Every production has an AI crew working in the background
  • Every delay is prevented before it happens
  • Filmmakers can focus on creativity, not logistics

That's the future we're building with Script-to-Screen Guardian.

🙏 Final Thanks

Built with ❤️ for the future of filmmaking.

Special thanks to:

  • Google Cloud for Gemini Enterprise and GEAR program access
  • ClickHouse for the free tier and excellent documentation
  • Devpost for organizing Agentic Cinema: The Blockbuster Hackathon
  • All 9,800+ participants for the inspiration and community support

🎬 Script-to-Screen Guardian | Predict. Prevent. Produce.

Built With

  • agent
  • agents
  • ai
  • artificial
  • builder
  • clickhouse
  • cloud
  • docker
  • enterprise
  • entertainment
  • fastapi
  • gemini
  • google
  • grafana
  • learning
  • machine
  • multi-agent
  • natural
  • python
  • streamlit
  • systems
Share this project:

Updates

Submission history