CinePulse AI: Autonomous Media Intelligence Agent

AI-powered studio intelligence system where Gemini autonomously queries ClickHouse OLAP to diagnose streaming drops, audit marketing ROAS, and generate executive directives all from natural language

Comment

Inspiration

Modern entertainment enterprises operate across massive, fragmented data streams — billions in box office receipts, millions of high-frequency streaming events, multi-million dollar marketing budgets split across TikTok, YouTube, Meta, and Connected TV. Yet studio executives still wait days for BI teams to write custom SQL reports.

We asked: What if an AI agent could autonomously investigate any strategic question by querying real-time OLAP data, diagnosing root causes, and formulating actionable directives — all in seconds?

That's CinePulse AI.

What it does

CinePulse AI is an autonomous media intelligence system for film studio executives and streaming platform leaders. It features:

  • 🤖 Gemini Director Agent — Ask any strategic question in natural language (e.g., "Why did Sci-Fi completion crash in Latin America?"). The agent autonomously plans a multi-step investigation, calls 12 analytical tools against ClickHouse, correlates data across tables, and synthesizes an executive dossier with evidence-backed directives.

  • 📊 Slate Performance Dashboard — Real-time genre financials (Box Office vs Budget vs Net Profit), global streaming distribution by territory, and multi-channel marketing ROAS attribution — all powered by live ClickHouse OLAP queries.

  • 🔍 ClickHouse OLAP Explorer — A SQL sandbox with pre-built query templates for territory completion rates, marketing spend attribution, monthly revenue trends, and platform anomalies. Sub-50ms query execution across 15,000+ events.

  • 🚨 Incident Control Room — Real-time observability monitoring for CDN packet loss, playback drop-offs, audio sync degradation, and retention anomalies — with one-click dispatch to the AI agent for autonomous investigation.

  • 💰 Capital Reallocation Simulator — Interactive slider-based tool that projects gross revenue lift, churn reduction, and net profit delta from reallocating media spend between channels.

  • ⚡ Deploy Directives — One-click deployment of AI-generated action directives with real-time status tracking (Idle → Executing → Deployed).

How we built it

Architecture:

  • Frontend: Next.js 15 (App Router) + React 19 + TypeScript + Tailwind CSS + Recharts for data visualization
  • AI Engine: Google Cloud Gemini API (gemini-2.5-flash) with multi-turn function calling — 12 registered analytical tools
  • Database: ClickHouse Cloud as the ultra-fast OLAP backbone, with a resilient dual-mode architecture (Cloud + embedded local engine for zero-friction offline execution)
  • Agent Pipeline: 4-stage autonomous investigation flow — PLANNING → QUERYING → ANALYZING → SYNTHESIZING — with full SQL audit trail

Key Technical Decisions:

  • Used Gemini's multi-turn tool calling to let the agent autonomously decide which ClickHouse tools to invoke based on the question
  • Designed 12 specialized analytical tools (revenue, retention, marketing ROAS, anomaly detection, content recommendations) as discrete function declarations
  • Built a resilient dual-mode ClickHouse connection — seamlessly falls back to embedded OLAP when cloud is unavailable
  • Dark cinematic "studio control room" UI design with premium glassmorphism aesthetics

Challenges we ran into

  • Multi-turn tool orchestration — Getting Gemini to chain multiple analytical tool calls in the right sequence required careful system prompt engineering and structured response schemas
  • ClickHouse schema design — Optimizing partition keys and sort orders for sub-50ms analytical queries across 15,000+ viewing events with multiple join patterns
  • Real-time chart generation — Translating raw ClickHouse aggregation results into dynamic Recharts visualizations with proper axis scaling and labeling
  • Resilient dual-mode architecture — Building a seamless fallback from ClickHouse Cloud to an embedded local engine without any code changes

Accomplishments that we're proud of

  • True autonomous agent behavior — The Gemini Director Agent formulates its own investigation plan, decides which tools to call, chains multiple ClickHouse queries, and synthesizes findings without any human guidance
  • 12 production-grade analytical tools — Each tool executes optimized parameterized SQL against ClickHouse with proper error handling and response normalization
  • Sub-50ms ClickHouse query latency — Achieved through careful schema design with MergeTree engines, partition pruning, and columnar compression
  • Premium cinematic UI — A dark studio control room aesthetic that feels like an enterprise-grade product, not a hackathon prototype

What we learned

  • Gemini's function calling is remarkably powerful for building autonomous agents — the model can reason about which tools to invoke and in what sequence
  • ClickHouse's columnar architecture makes it ideal for the kind of multi-dimensional analytical queries that media intelligence requires
  • System prompt engineering is critical — the quality of the agent's investigation and recommendations improved dramatically with structured persona constraints and evidence-quotation rules

What's next for CinePulse AI

  • Real-time streaming ingestion — Connect to Apache Kafka for live viewing event streams into ClickHouse
  • Gemini 2.5 Pro upgrade — Leverage longer context windows for more complex multi-step investigations
  • Collaborative investigations — Multiple executives collaborating on the same investigation with shared annotations
  • Predictive analytics — Use historical ClickHouse data to predict churn, forecast box office performance, and recommend optimal release windows
  • Google Cloud Run deployment — Containerized production deployment with auto-scaling

Built With

Share this project:

Updates

Submission history