The problem
A production change rarely stays isolated. A Scene 12 wardrobe switch from a blue jacket to a black jacket can conflict with footage already shot, change the next call sheet, and require Wardrobe and the Assistant Director to act quickly. The evidence exists, but it is scattered across production records and schedules.
The solution
SlateGuard is an evidence-first continuity command desk for that decision. A script supervisor starts a bounded production revision. SlateGuard records the change, retrieves the current evidence window, identifies continuity and scheduling impact, produces a grounded Change Packet, and waits for a human to create the follow-up.
The core loop is intentionally clear and inspectable:
revision → ClickHouse writer MCP → ClickHouse reader MCP evidence → Gemini Change Packet → human decision → writer MCP → reader-verified receipt
Why ClickHouse matters
ClickHouse is SlateGuard's time-aware agent memory, not background storage. Fixed reader-MCP queries return only the established Scene 11 history, active Scene 12 revision, and next scheduled dependencies. Unrelated, stale, and unscheduled records stay outside the agent context.
Reader and writer use separate, least-privilege paths. The browser and model receive neither raw database credentials nor arbitrary SQL. The writer appends named production events; the reader is limited to curated evidence views.
AI and human authority
Google ADK and Gemini produce a structured Change Packet from bounded facts, approved evidence IDs, deterministic findings, and allowed owners. Invalid citations, unsupported status changes, and unapproved owners are rejected. Gemini explains the evidence; it does not decide a consequential production action.
A human supervisor creates the recommended Wardrobe and Assistant Director follow-up. SlateGuard then independently reads the result back through the reader path before changing the visible readiness state from At risk to Follow-up created.
What the live demo proves
The deployed Cloud Run app shows the full Scene 12 wardrobe loop using six self-authored fictional production scenes:
- official ClickHouse MCP protected write and curated reader retrieval;
- a visible current-evidence count and runtime proof;
- a Gemini-grounded, schema-constrained Change Packet;
- human-owned follow-up creation; and
- a reader-verified readiness receipt.
Built with
React, TypeScript, Vite, Python, FastAPI, Google Cloud Run, Vertex AI, Google ADK, Gemini, ClickHouse Cloud, the official Python mcp-clickhouse server, Docker, Cloud Build, Artifact Registry, Secret Manager, and GitHub.
Testing
The project includes 43 backend tests covering API contracts, the fixed ClickHouse MCP boundary, production-memory parsing, session/idempotency behavior, and Change Packet validation. The public app is designed to be exercised as a complete revision-to-receipt loop.
Log in or sign up for Devpost to join the conversation.