ProductionPulse

Autonomous AI film studio analytics & multi-agent production crew uniting Google Gemini 2.5 Pro and ClickHouse Cloud to transform screenplays into set-ready logistics in sub-milliseconds.

Comment

Inspiration

Film and high-end episodic television production is a multi-million-dollar logistical puzzle executed under extreme time pressure. A standard 120-page screenplay takes line producers, 1st ADs, and unit production managers days of manual labor to break down scenes, calculate department budgets, and schedule shooting blocks.

Unoptimized shooting schedules lead to massive budget leaks—most notably SAG-AFTRA actor "holding penalties," where principal cast must be paid full day rates for idle days between active scenes. Furthermore, studio executives lack the tools for real-time financial forecasting. When weather delays or location changes hit, answering "What happens to our bottom line if we cut two night exterior shoot days?" requires hours of spreadsheet re-work.

We built ProductionPulse to give studios and independent filmmakers an autonomous AI production crew that parses screenplays into structured data and executes complex logistical analytics in sub-milliseconds using ClickHouse Cloud and Google Gemini 2.5 Pro.

What it does

ProductionPulse transforms raw screenplay files into a comprehensive, interactive studio intelligence platform:

1) Multimodal Screenplay Ingestion: Ingests PDF, TXT, or Fountain screenplays, automatically extracting structured scenes, cast appearances, stunts, props, VFX shots, and complexity scores. 2) Executive Overview Dashboard: Real-time financial breakdowns, department allocations, shooting day estimates, and location splits. 3) Autonomous AI Studio Production Crew: A collaborative multi-agent network (Line Producer, 1st AD, SAG-AFTRA Talent Coordinator, Safety & VFX Supervisor) communicating over the Model Context Protocol (MCP). 4) Live ClickHouse MCP SQL Inspector: Plain-English queries are translated by Gemini into optimized ClickHouse SQL, executed over persistent stdio MCP sessions with sub-millisecond OLAP speed (<2ms query execution, ~500ms full end-to-end agent reasoning), and displayed transparently in real-time. 5) SAG-AFTRA Day-Out-of-Days (DOOD) Matrix: Color-codes Start (SW), Work (W), Hold (H), and Drop (WF) states across the shooting calendar to eliminate idle holding fees. 6) Hollywood Stripboard Scheduler: Color-codes scene strips by DGA industry lighting conventions (White Int Day, Amber Ext Day, Green Int Night, Blue Ext Night) grouped by location to minimize costly company moves. 7) Real-Time What-If Scenario Simulator: Interactive sliders allow producers to simulate VFX cutbacks, daily shooting caps, and salary adjustments with instant recalculation. 8) Exportable Industry Call Sheets: Generates production-ready daily call sheets with cast schedules, Leavesden Stage locations, meal breaks, and hospital emergency routing. 9) Zero-Config Judge Demos: Instant 1-click datasets (Mind Heist & Cyber Horizon) for immediate evaluation without requiring uploads.

How we built it

  • ClickHouse Cloud & mcp-clickhouse: Serves as the core analytical backbone. Production data is stored in columnar ClickHouse MergeTree tables (scenes, characters, locations, budget_items, schedule_items) with real-time materialized views (scene_summary). Our backend connects via a persistent StdioServerParameters client session to the official mcp-clickhouse server over JSON-RPC stdio transport.
  • Google Gemini 2.5 Pro & Google ADK: Powers screenplay multimodal extraction, multi-agent persona reasoning, and natural-language-to-SQL translation.
  • FastAPI Backend (Python 3.11): Orchestrates agent protocols, rolling latency percentiles (p50: 557ms), and research grounding.
  • Next.js 14 & TailwindCSS Frontend: Cinematic dark-mode dashboard featuring glassmorphism, Recharts analytics, and Radix UI components deployed to Firebase Hosting.

Challenges we ran into

1) True MCP Architecture vs Direct DB Calls: The contest rules require active integration with the Model Context Protocol. We re-architected our backend from direct in-process drivers into a persistent stdio subprocess server communicating via formal JSON-RPC ClientSession.call_tool() protocols. 2) Reconciling Sub-Millisecond OLAP vs Network Latency: ClickHouse Cloud processes aggregations server-side in <2ms. Managing the full round trip (Gemini 2.5 Pro token generation + MCP JSON-RPC framing + Asia-Southeast1 network hops) down to a stable ~500ms p50 required persistent connection pooling. 3) SAG-AFTRA Union Hold Logic: Translating complex collective bargaining rules (consecutive work rules, drop-and-pickup thresholds) into deterministic algorithmic matrix generators.

Accomplishments that we're proud of

  • Sub-millisecond server-side ClickHouse analytical queries powering real-time sliders. Full transparency with the live MCP SQL Inspector showing judges the exact query executed on ClickHouse Cloud.
  • Zero-config judge demo dataset launcher that works out of the box with zero setup.
  • Hollywood-standard DOOD matrix that calculates real dollar savings on holding day penalties.

What we learned

Combining high-throughput columnar databases like ClickHouse with agentic LLMs unlocks entirely new categories of operational intelligence for creative industries that have traditionally relied on static paper binders and slow spreadsheets.

What's next for ProductionPulse

  • Real-time on-set IoT telemetry integration for live budget burn-rate tracking.
  • Autonomous multi-camera shot list generation and DGA digital script supervisor sync.

Built With

Share this project:

Updates

Submission history