💡 What inspired us Continuity errors cost Hollywood millions of dollars in reshoots, CGI fixes, and post-production delays (we all remember the infamous coffee cup in Game of Thrones). Keeping track of thousands of props, wardrobe states, character locations, and lighting setups across non-linear shooting schedules is a complete nightmare for on-set Script Supervisors.
We realized that modern AI could solve this, but LLMs alone hallucinate too much for strict continuity tracking. We needed an intelligent, immutable, and blazingly fast memory layer for film production sets that can deterministically flag inconsistencies in real-time. That's how CINESTATE was born.
🎬 What it does CINESTATE is an autonomous Agentic Media Production State & Continuity Intelligence Service.
It ingests script breakdowns and live observations from the set. An autonomous Orchestrator Agent (powered by Google ADK) actively monitors incoming takes. When it detects a discrepancy between the established "truth" in the database and a new real-time observation (e.g., John's tie was Red in Scene 2, but it's Blue in Scene 4), it instantly:
Flags a Continuity Conflict. Calculates the Blast Radius (which future scenes are affected by this mistake). Provides Actionable Recommendations to the director before the set is struck and the actors go home.
⚙️ How we built it We architected a robust, production-grade system: Frontend Command Center: Built with React, Vite, and TailwindCSS to provide a beautiful, real-time dashboard for directors and script supervisors. AI Service Core: A Python FastAPI backend powered by the Google Agent Development Kit (ADK) and the Google Gemini Pro model. The Brain & Memory: We used ClickHouse Cloud as the immutable state machine and event ledger. Its columnar speed allows our agents to query thousands of historical production events in milliseconds. MCP Integration: We utilized the Model Context Protocol (MCP) via the official mcp-clickhouse server. This allows our ADK Orchestrator to securely and dynamically query ClickHouse to establish ground truth, ensuring the AI relies on hard data rather than hallucinations.
🚧 Challenges we ran into
Bridging the Deterministic and the Generative: We had to carefully prompt engineer our ADK Orchestrator to never decide a continuity conflict exists from language alone. It must always retrieve historical state from ClickHouse via MCP tools first, and only flag a conflict if a deterministic mismatch occurs. Temporal State Logic: Designing a database schema in ClickHouse that could handle temporal state changes (e.g., A character gets a scar in Scene 5, so Scenes 6-10 must have the scar, but Scenes 1-4 must not, regardless of the order they are filmed in).
🏆 Accomplishments that we're proud of
Building a genuinely useful Agentic workflow that solves a real-world, multi-million dollar industry problem. Crafting a premium, Hollywood-grade UI/UX that feels like a native production tool. Successfully integrating ClickHouse Cloud as the primary, real-time "Memory" for an autonomous multi-agent system.
📚 What we learned
We learned how incredibly powerful the Google ADK is for orchestrating multiple specialized sub-agents (Evidence Agent, Script Agent, Recommendation Agent) under one cohesive workflow. We discovered that ClickHouse Cloud is the perfect companion for LLMs. Its extreme query speed enables real-time AI decision-making without adding unacceptable latency to the fast-paced environment of a film set.
🚀 What's next for CINESTATE
Direct API integration with live camera feeds from ARRI and RED digital cinema cameras. Multi-modal support for audio continuity (matching ambient room tone and actor cadence across takes). A native iPad companion app for script supervisors to use directly on set.
Log in or sign up for Devpost to join the conversation.