Where CineSpine uses AI

Where CineSpine uses AI An architecture diagram generated by Archify. WebMCP tools · document.modelContext, 8 tools · Architecture component WebMCP tools document.modelContext, 8 tools Production status · GET /productions/{id}/status · cinespine backend (Cloud Run) Production status GET /productions/{id}/status Ingestion parsers · classifier, report vision · cinespine backend (Cloud Run) Ingestion parsers classifier, report vision Script Studio AI · profiles, shots, images · cinespine backend (Cloud Run) Script Studio AI profiles, shots, images ADK agents · Wrap Rescue, AE Queue · cinespine backend (Cloud Run) ADK agents Wrap Rescue, AE Queue llm_router · 4-model fallback + cache · cinespine backend (Cloud Run) llm_router 4-model fallback + cache cinespine-mcp · mcp-clickhouse, read-only · cinespine backend (Cloud Run) cinespine-mcp mcp-clickhouse, read-only Google Cloud AI APIs · Document AI, Speech, Video · Architecture component · Google Cloud Google Cloud AI APIs Document AI, Speech, Video Gemini · text + image models · Architecture component · Google Gemini Gemini text + image models ClickHouse Cloud · analytical mirror · Architecture component · ClickHouse ClickHouse Cloud analytical mirror Grafana Cloud · GenAI spans, tokens · Architecture component · Grafana Grafana Cloud GenAI spans, tokens set DoP, render, select run Wrap Rescue summarize_production_status OCR, transcripts, shot detection classify, read reports JSON prompts LlmAgent model generate_content list_tables, run_query SELECT only GenAI instrumentation cinespine backend (Cloud Run)

Generative uses

  • • Character profiles, shot breakdowns and DoP preset copy are Gemini Flash JSON with a deterministic fallback when the model is unavailable
  • • Portraits and storyboards try three Gemini image models, then serve a labelled placeholder rather than a fake render

Reading, not writing

  • • Uploads are classified by Gemini Flash plus rules; camera reports go through vision; Document AI, Speech and Video Intelligence extract text from scans, audio and footage
  • • The discrepancy assistant explains a take from spine facts; it never edits them

Agents and guardrails

  • • Wrap Rescue is an ADK LlmAgent that reaches ClickHouse only through cinespine-mcp; the AE Queue agent is deterministic and uses no model
  • • summarize_production_status gives an agent one ranked answer (done, running, blocking, left, missing; severity then age); llm_router walks four Gemini models on 503 and caches by prompt hash