MISE

AI Production Memory for filmmakers, crews, and studios.

Comment

Inspiration

Film productions generate an enormous amount of knowledge: scene decisions, approved assets, camera choices, lighting changes, continuity notes, production events, and creative reasoning.

The problem is that this knowledge rarely lives in one place.

A director may remember why a 35mm lens was chosen. An editor may know which asset version was approved. A continuity team may have a critical note about wardrobe or story logic. But as production grows, those decisions become scattered across conversations, documents, folders, tools, and individual memories.

That inspired MISE.

MISE is an AI-powered Production Memory system designed to help filmmakers, producers, editors, VFX teams, and studio crews retrieve trusted production knowledge instantly.

Instead of asking:

“Who remembers why we changed the lighting in Scene 41?”

a production team can simply ask MISE.

MISE turns fragmented production history into a queryable, persistent memory layer for filmmaking.

What it does

MISE creates a searchable production-memory layer for film and media workflows.

Users select a production and scene, then ask natural-language questions about:

  • approved assets
  • scene history
  • camera decisions
  • lighting decisions
  • continuity
  • creative reasoning
  • production events

For the hackathon demo, MISE uses the fictional production SCOPE CREEP.

Example questions include:

  • Which version of Lauren’s entrance master is approved?
  • What camera decision should editorial preserve for Scene 23?
  • Why was the lighting changed in Scene 41?
  • What continuity note should the team remember for Scene 41?

MISE retrieves the relevant production records at runtime and returns a concise, grounded answer instead of inventing production information.

For example, MISE can retrieve that:

  • Lauren entrance master v7 is the approved asset for Scene 23
  • editorial should preserve the decision to use a 35mm lens for Lauren’s entrance
  • the reason was to keep the smart-home environment visible while preserving intimacy
  • Scene 41 was shifted toward colder moonlight to increase tension
  • the house should show a subtle preference for Lauren before the explicit reveal

The long-term product model is multi-production: production teams create or select a project, add production materials, and use MISE as a living memory layer throughout development, production, post-production, and VFX.

How we built it

ClickHouse runtime verification: MISE actively uses the official mcp-clickhouse MCP server at runtime. The public repository includes the deployed backend implementation in backend/app/agent.py, backend/app/tools.py, backend/Dockerfile, and backend/mcp_cloud_test.py. The Dockerfile installs mcp-clickhouse==0.6.0, and Google ADK connects to the ClickHouse MCP server through McpToolset and StdioConnectionParams. MISE combines Google Cloud agent infrastructure with ClickHouse as its production-memory layer.

Google Cloud AI and agent layer

The agent is built with:

  • Gemini 2.5 Flash
  • Vertex AI
  • Google Agent Development Kit (ADK)
  • Google Cloud Run

Google ADK manages the agent and tool-calling workflow.

Gemini interprets the filmmaker’s question, identifies the relevant production-memory task, and selects the appropriate tool.

The agent does not rely on a static answer set. Production information is retrieved dynamically at runtime.

ClickHouse production-memory layer

Structured production data is stored in ClickHouse Cloud.

The demo database contains production-memory tables representing:

  • productions
  • scenes
  • assets
  • decisions
  • continuity
  • production events

MISE actively integrates ClickHouse at runtime through the official mcp-clickhouse MCP server, connected to a ClickHouse Cloud cluster.

The ClickHouse integration is part of the live application workflow, not only part of development or documentation.

The runtime architecture is:

User → MISE UI → Cloud Run → Google ADK → Gemini on Vertex AI → MISE production-memory tools → official mcp-clickhouse MCP server → ClickHouse Cloud → grounded response

The MCP connection exposes ClickHouse operations including:

  • list_databases
  • list_tables
  • run_query

These capabilities allow MISE to inspect and query the production-memory database during execution.

Semantic production-memory tools

Rather than asking Gemini to generate unrestricted SQL for every request, we created semantic tools for common filmmaking memory tasks.

These include:

  • resolve_scene
  • get_approved_assets
  • get_scene_decisions
  • get_scene_continuity

The semantic tools translate production questions into controlled ClickHouse queries executed through the official ClickHouse MCP runtime.

For example:

get_scene_decisions(
    scene_number="23",
    category="camera decision"
)

Built With

  • ai
  • clickhouse
  • clickhouse-cloud
  • docker
  • gemini-2.5-flash
  • google
  • google-agent-development-kit-(adk)
  • google-cloud-run
  • official-mcp-clickhouse-mcp-server
  • python
  • secret-manager
  • vertex-ai
Share this project:

Updates

Submission history