RenderOps Director helps a render supervisor understand why a cinematic shot failed and recover only the affected frames.
Live demo: https://renderops-director-w6mw3t2ita-uc.a.run.app
Hackathon: Agentic Cinema — Grafana track
A failed render is more than a red infrastructure alert. The supervisor needs to know which frames failed, what changed, whether the deadline is at risk, and whether a full rerender is really necessary.
The demo follows shot SH-042. A texture update pushes GPU memory to 96%, the final denoise pass starts failing, and 38 of 240 frames become unusable.
- Compare Original plate and Failed render.
- Click Investigate shot.
- Review the evidence from Prometheus, Loki, and Tempo.
- Approve a five-frame canary.
- Wait for Grafana to validate the new render.
- Approve recovery of the 38 failed frames.
- Watch the generated Recovered shot.
The recommended path costs $35.90 instead of $186.40 for a full rerender.
The app sends render telemetry to Grafana Cloud through OpenTelemetry. A Gemini agent running on Google Cloud uses the official Grafana MCP server to query:
- Prometheus for frame failures, GPU pressure, queue delay, and cost;
- Loki for renderer and FFmpeg logs;
- Tempo for the failed denoise path and recovery traces.
The agent turns those signals into a recovery plan. Grafana is queried again after every approved step, so recovery does not unlock until the new telemetry passes its checks.
The initial GPU out-of-memory incident is a controlled synthetic scenario. The approved work is real:
- FFmpeg runs inside Cloud Run;
- canary and recovery requests create new WebM files;
- the browser plays the file returned by the current request;
- exit code, duration, frame count, output size, logs, and traces are sent to Grafana;
- Grafana MCP reads that evidence back before the workflow continues.
No Grafana write tools are exposed. Expensive actions require human approval.
The editable diagram is available as SVG. A more detailed evidence map is in docs/grafana-evidence.md.
Requirements: Python 3.11+, Docker, Google Cloud credentials, and optionally Grafana Cloud credentials for live mode.
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
uvicorn app.main:app --reloadOpen http://localhost:8000.
The default mode uses local demo data. Live settings are documented in .env.example. Tokens belong in environment variables or Google Secret Manager, never in the repository.
python -m ruff check .
python -m pytest
python -m compileall app
docker build -t renderops-director:test .Google ADK, Gemini on Vertex AI, Grafana Cloud, official mcp-grafana, Prometheus, Loki, Tempo, OpenTelemetry, Cloud Run, FastAPI, FFmpeg, and Playwright.
MIT © 2026 Aleksei Chirkunov