ReelLedger

Live financial exposure and comparable-title intelligence for productions that can't afford Cinelytic — Gemini agents querying ClickHouse via MCP, in plain English.

Comment

Inspiration

Two things are true about film & TV production right now, and they compound. First, production finance is still fragmented across spreadsheets — industry coverage on production accounting is unanimous that studios want live budget visibility, not a month-end reconciliation, but most productions below the studio tentpole tier are still stitching this together by hand. By the time a department head notices they're 20% over on a line item, the money's already spent. Second, predictive "will this get made / will this perform" analytics — tools like Cinelytic — exist, but they're priced and built for major studios. Indie and mid-budget productions, who most need to know whether a project is financially viable before they commit, have no equivalent. We found an industry software directory that explicitly separates "studio-grade software that uses intelligent analysis to boost financing and greenlight chances" from the commodity budgeting tools everyone else uses — that gap is what ReelLedger fills.

What it does

ReelLedger is a small multi-agent system a producer talks to in plain English. Ask "which department is trending over budget" and the Exposure Agent queries live spend data and tells you, with numbers, which department and by how much. Ask "we're an $8M elevated horror film with an unknown lead releasing Q4 — what should we expect" and the Comps Agent pulls comparable historical titles and gives you a distribution (median, P25, P75), not a single made-up number. A Producer Agent routes the question and can combine both — e.g. "if VFX keeps overspending at this rate, are comparable films at our new effective budget still viable?"

How we built it

Built end-to-end inside Google Antigravity, Google's agentic development platform — fitting, since the app itself is an agentic system. Google Cloud Agent Builder (ADK) + Gemini orchestrate the three agents. All analytical data — production spend line items and 2,000 synthetic comparable-title records — lives in ClickHouse Cloud, queried by the agents at runtime through the official ClickHouse MCP server (mcp-clickhouse), not a hardcoded query layer. We chose ClickHouse specifically because both data shapes here — high-cardinality time-series spend events and large comparable-title tables — are exactly the columnar-aggregation workload it's built for; a producer's ad hoc natural-language question needs sub-second response even against thousands of line items. The app is deployed on Cloud Run, with Vertex AI handling Gemini inference (no API keys in the deployed service — IAM-scoped service account) and ClickHouse credentials in Secret Manager.

Challenges we ran into

The ClickHouse MCP server's actual tool name (run_query) didn't match what we initially assumed (run_select_query) — worth checking directly against the server rather than trusting docs/memory. We also found that spawning the MCP server via uvx at runtime meant a PyPI fetch on every cold start in Cloud Run, which is slow and adds an external dependency at request time; we switched to spawning it directly via the active Python interpreter (sys.executable) so the container is fully self-contained after build.

Accomplishments that we're proud of

Real, non-mocked ClickHouse-via-MCP integration exercised at runtime — every number the agents state comes from a live query, not a cached or invented figure. And a demo dataset seeded with a genuine cost-overrun pattern (VFX running ~20% hot against schedule progress) so the Exposure Agent's finding is a real signal, not a scripted answer.

What we learned

How ADK's MCPToolset wiring actually behaves in production versus docs (tool naming, subprocess lifecycle), and how much of the "enterprise analytics for indie film" gap is validated by the industry's own trade coverage — this isn't a hypothetical problem.

What's next for ReelLedger

Multi-project support (right now the demo seeds one production), a write path so department heads can log spend directly instead of a nightly import, and swapping the synthetic comps dataset for a licensed real-world data source.

Built With

Share this project:

Updates

Submission history