CineCompute

Inspiration

Modern VFX and animation studios (Houdini Karma, Maya Arnold, Unreal Engine, Nuke) suffer from catastrophic cloud compute overruns. High-end GPU nodes crash silently due to out-of-memory (OOM) leaks from dense volumetric simulations, or get trapped in expensive retry loops. Technical Directors and VFX Producers typically discover these failures weeks later on cloud invoices, after tens of thousands of dollars have evaporated. We built CineCompute to give production supervisors real-time, conversational observability and prescriptive FinOps automation over their render farm telemetry.

What it does

CineCompute is an autonomous FinOps intelligence engine that transforms raw telemetry into actionable pipeline decisions:

• Sub-Millisecond Telemetry Querying: Ingests and analyzes 250,000+ render events across multi-shot projects in real time.
• Root-Cause Anomaly Detection: Pinpoints silent OOM failures, driver crashes on specific GPU pools (e.g., L40S, A100), and unoptimized shaders with shot-level precision.
• Predictive Budget Forecasting: Extrapolates true project completion costs by joining delivered frames against target frames and current burn rates, flagging overruns weeks before delivery deadlines.
• Artist-Level Attribution & Coaching: Identifies disproportionate failure rates per artist without blame, automatically generating constructive Slack remediation drafts with scene optimization tips.
• Instant Farm Audits & Scheduler Policies: Sweeps the entire farm via analytical SQL to generate downloadable, production-ready JSON scheduler rules for AWS Thinkbox Deadline and Tractor.

How we built it

• Intelligence Layer: Powered by Google Gemini (Gemini 3.1 Pro / 3.8 Flash) using the official google-genai SDK for multi-step reasoning, dynamic schema introspection, and automated arithmetic verification.
• High-Speed OLAP Engine: Powered by ClickHouse Cloud (GCP deployment) storing telemetry in optimized MergeTree tables with sub-20ms aggregate performance.
• Model Context Protocol (MCP): Integrated via the official mcp-clickhouse server running as a subprocess over stdio, enforcing readonly=1 safety sessions and an allowlist of analytical read queries.
• Observability UI: Built with Streamlit using a high-density, low-ink observability design system (inspired by Stripe and Linear) with interactive Plotly charts and a live MCP Query Inspector displaying raw SQL, execution latency, and row counts.

Challenges we ran into

Preventing LLM hallucinations on mission-critical financial figures was our biggest challenge. We solved this with a two-fold defense:

1. Enforcing strict SQL syntax and semantic rules in the prompt (e.g., distinguishing total spend from waste via sumIf, forbidding unchecked percentages).
2. Implementing an automated arithmetic verification agent that cross-checks the LLM's derived figures against direct ClickHouse ground-truth data before presenting them.

Accomplishments that we're proud of

• Enabling sub-20ms analytics over 250,000 events via ClickHouse Cloud.
• Delivering a zero-mock, fully inspectable MCP pipeline where every single SQL query and its latency are transparently displayed to the user.
• Moving from passive observability to closed-loop remediation: generating concrete scheduler policies and artist-facing action plans from raw telemetry.
• 100% Google Cloud and ClickHouse native architecture with zero third-party AI dependencies.

What we learned

Pairing ClickHouse's blazing-fast OLAP capabilities with the autonomous tool-calling precision of Google Gemini creates a new standard for industrial FinOps. When an agent is given direct, secure access to analytical primitives via MCP, it ceases to be a generic chatbot and becomes an indispensable pipeline engineer.

What's next for CineCompute

• Native bi-directional API plugins for AWS Thinkbox Deadline and OpenCue to automate real-time job quarantine and high-VRAM node re-routing.
• Multimodal inspection using Gemini to visually detect corrupted or NaN-pixel output frames before they reach compositing.

Built With

• clickhouse
• docker
• google-cloud
• google-gemini
• mcp
• pandas
• plotly
• python
• streamlit