Film Pipeline: typed state for AI-assisted film production

github.com/ghassan-ai-projects/film-pipeline-langgraph — Python, LangGraph, MCP, Textual, MIT licensed.

Film Pipeline takes an idea through structured pre-production, generation planning, QC evidence, and validated handoff state. It is not final editing software, and it does not claim to generate a finished film by itself. It is the orchestration and evidence layer before downstream generation and finishing.

The supported product target is deliberately narrow:

idea -> approved artifacts -> generation planning -> QC-ready project state

The project treats creative work as stateful production, not a sequence of disconnected prompts. Artifacts are versioned. Approvals are explicit. Repairs preserve dependency history. Provider decisions and validation reports can be inspected.

Why I built it

AI film workflows often fail because the process is invisible. Chat history becomes the production system. That does not scale when a screenplay, film matrix, shot bible, reference plan, generation batch, cost approval, provider failure, or QC issue needs to survive revisions and handoffs.

The project grew out of real pipeline failures: moderation blocks, bad cost assumptions, broken sequential generation chains, missing checkpoints, and validators that were too close to the work they reviewed. The rebuild turns those failures into graph state, gates, artifacts, and recovery paths.

What to inspect

• Typed LangGraph state contract with append-only reducers.
• Human approval gates using real LangGraph interrupts and resume commands.
• Versioned artifacts with dependency tracking.
• Generation planning with explicit spend approval and provider routing.
• Parallel QC validators using LangGraph fan-out.
• Terminal operator console, headless CLI, and MCP tools over the same state.

Interfaces

The terminal studio runs with uv run film-pipeline-tui. The headless runner uses uv run film-pipeline-run my-idea.txt. The MCP server exposes project, artifact, review, validation, generation, provider, checkpoint, rollback, audit, and knowledge-base tools for agent operation.

Current limits

Mock mode verifies the machinery, not final model quality. Real-provider mode can spend money and depends on configured provider keys. Post-production is planned, not fully executed inside the repository: final editing, sound, color, and export remain outside the current product target.

Why it belongs here

This project shows the same architecture discipline in a creative domain: state contracts, checkpoints, validation, rollback, cost boundaries, and inspectability. The medium is film creation, but the engineering problem is reliable agent workflow design.

Read the open-source release article or review the original postmortem.