CineOps Guard

Autonomous film production OS automating script intake, safety compliance, pre-vis storyboarding, and video dailies hazard inspection.

Comment

🎬 CineOps Guard: Multimodal AI Dailies & Safety Compliance Suite

Inspiration

High-stakes film production is creative, dynamic, and inherently dangerous. Analyzing daily set footage ("dailies") for stunt hazards, OSHA violations, and SAG-AFTRA compliance traditionally requires hours of manual review by stunt coordinators and script supervisors. A single missed safety protocol can lead to severe injury, production halts, and budget overruns.

We asked: How can AI act as a creative co-pilot without compromising set safety? Using Gemini 3.6 Flash, Google ADK, and GCP, we built an agentic co-pilot that provides real-time visual hazard detection, quantitative risk scoring, and automated compliance logging.

What it does

CineOps Guard is a multimodal multi-agent platform for film production safety and dailies auditing:

  1. Automated Visual Dailies Auditing: Analyzes raw footage, transcribes dialogue, extracts key visual action frames, and flags safety hazards.
  2. Deterministic Safety Guardrails: Calculates a quantitative risk score $S_{\text{risk}}$ against safety rules. If $S_{\text{risk}} \ge 0.70$, scene sign-off is blocked until a human safety officer provides an override.
  3. Live Rehearsal API: Provides dynamic, real-time voice coaching sessions for actors via WebSockets.
  4. Pre-Visualization Generation: Creates production storyboards (Imagen 3) and mood audio tracks (Lyria 3) directly from script parsing.
  5. MCP Compliance Engine: Serves enterprise audit logging to Google BigQuery and GCS.

How we built it

  • Multi-Agent Orchestration: Built with Google ADK featuring a root DirectorAgent delegating to 5 specialized sub-agents (DailiesAgent, ComplianceAgent, ScriptIntakeAgent, StoryboardAgent, VoiceRehearsalAgent).
  • Deterministic Risk Model:

$$S_{\text{risk}} = \min \left( 1.0, \sum_{i=1}^{n} w_i \cdot V_i + \gamma \cdot (1 - C_{\text{harness}}) \right)$$

Where $V_i$ are detected hazard flags, $w_i$ are weights, and $C_{\text{harness}}$ is harness detection confidence.

  • Model Context Protocol (MCP): Standardized tool-execution server interfacing with BigQuery and GCS.
  • Full Stack & Cloud Infrastructure: Next.js 14 frontend, FastAPI API Gateway, containerized with Docker, deployed to GCP Cloud Run and Vertex AI Agent Engine via automated GitHub Actions CI/CD.

Challenges we ran into

  • Cloud Run Cold Starts & Port Binding: Heavy ADK imports at startup caused port-binding timeouts. We resolved this by lazy-loading LLM runners inside FastAPI's lifespan handler and configuring shell-form ${PORT} execution in Docker.
  • IAM Log-Streaming in Cloud Build: Service account permission constraints broke gcloud builds submit. We re-architected the CI/CD pipeline to use direct docker build + docker push to GCR within GitHub Actions.
  • Strict ADK Schemas: Avoided Pydantic validation errors by configuring types.SafetySetting guardrails directly on generation calls rather than the root agent constructor.

Accomplishments that we're proud of

  • One-Click Cloud Deployment: Built a production-grade CI/CD pipeline deploying 6 multi-modal agents seamlessly to GCP.
  • Hybrid Intelligence: Combined non-deterministic LLM reasoning with deterministic mathematical safety guardrails for zero-false-negative stunt safety checks.
  • Sub-Second Audio & Vision Pipeline: Achieved real-time visual stunt hazard detection and WebSocket voice coaching.

What we learned

  1. Lazy Loading for Serverless AI: Deferring heavy model runners until after web server port binding is critical for serverless container stability.
  2. Model Context Protocol (MCP): Decoupling tools from agent logic allows instant swapping between local dev databases and GCP BigQuery without code changes.

What's next for CineOps Guard

  • Edge Live Streaming: Connecting live RTSP camera feeds on physical film sets for instant real-time hazard alerts.
  • SAG-AFTRA Contract Automation: Auto-auditing union rest periods, meal penalties, and minor working hours.
  • 3D Hazard Mapping: Utilizing Gemini 3.6 spatial bounding boxes to calculate physical distances between pyrotechnics and crew members.

Built With

Share this project:

Updates

Submission history