SmartCut Analytics

Home page
Diagnostic report – Part 1
Diagnostic report – Part 2

Inspiration

In the Media & Entertainment (M&E) industry, every second of a movie trailer counts. A single pacing drop or cognitive friction point can cause massive audience drop-offs, costing studios millions of dollars in potential box office revenue. Traditionally, identifying where and why viewers lose interest is a slow, manual, and subjective forensic process. We were inspired to bridge the gap between hard data and creative reasoning by building an autonomous "StreamOps" platform that acts as an elite Post-Production consultant in real-time.

What it does

SmartCut Analytics is an AI-powered Streaming Operations agent that proactively optimizes video audience retention.

Instead of just showing dead data, the platform calculates the drop-off severity and triggers an autonomous workflow:

Ingests massive amounts of real-time viewer telemetry.
Deterministically detects the exact second where the highest audience churn occurs.
Orchestrates an AI Agent to cross the critical drop-off timestamp with the scene's technical script, generating a professional post-production memorandum.

To determine the severity of a drop-off, our system relies on continuous retention monitoring, which can be mathematically modeled as:

$$ \text{Retention}(t) = \left( 1 - \frac{\sum_{i=0}^{t} \text{Dropoffs}_i}{\text{Total Viewers}} \right) \times 100 $$

When $ \text{Dropoffs}_t $ exceeds the standard threshold, the Gemini Agent is deployed to fix the scene.

How we built it

We built a production-grade, end-to-end enterprise stack:

Real-time Analytics: We used ClickHouse running locally on Docker. Its MergeTree engine allowed us to ingest and query thousands of simulated viewer events in milliseconds, serving as the perfect high-performance trigger.
Agentic Reasoning: We integrated Google's Generative AI SDK, specifically leveraging the Gemini 3.6-Flash model for rapid, quota-friendly inference.
Frontend & Control: We developed a resilient, neon-styled cinematic UI. We implemented session state controls to handle asynchronous AI calls gracefully.

Here is a quick look at how we route the context dynamically to the agent:

# Agentic reasoning implementation
prompt = f"Analyze the telemetry alert: Drop-off at Second {crit_sec} during the scene: '{escena_critica}'. Generate a creative diagnosis, a re-editing proposal, and the expected impact."

model = genai.GenerativeModel('gemini-3.6-flash')
response = model.generate_content(prompt)

Built With

agentic-ai
ai-agents
big-data
clickhouse
data-engineering
data-science
docker
ecosystem
ecosystem-integration
enterprise-workflows
gemini-api
generative-ai
google-cloud
machine-learning
media-and-entertainment
open-source
plotly
post-production
python
real-time-analytics
sql
streamlit
streamops
video-analytics

Submitted to

Agentic Cinema: The Blockbuster Hackathon

Created by

As the sole developer on this project, I was responsible for the entire end-to-end architecture, from system design to final deployment. This was my first time orchestrating a full, real-time data pipeline connecting a local, containerized database with a cloud-based generative AI.

My main contributions were:

System Architecture: I designed the complete workflow, deciding to use ClickHouse on Docker for its high-performance analytical capabilities, which was crucial for real-time telemetry ingestion.

Back-End Development: I wrote all the Python code to connect the Streamlit frontend with the ClickHouse database, implementing deterministic SQL queries to detect audience drop-off anomalies.

AI Agent Engineering: I was responsible for the prompt engineering and integration of the Gemini API. I learned a lot about handling API errors, such as rate limits (429) and model versioning (404), which led me to architect a more resilient system that intelligently routes requests to the latest available models like gemini-1.5-flash.

Front-End and UX: I designed and built the entire interactive dashboard using Streamlit and Plotly, including custom CSS for the cinematic feel and implementing session-state logic to create a non-blocking, user-friendly experience during AI processing.

It was an amazing learning experience, especially in understanding that building powerful AI agents is less about writing simple code and more about orchestrating the complex flow of data between high-performance systems.

Jorge Armando Alpala Aguilar