It's time for Industry BrickTalk #3 - Turning video into intelligence at scale.
In this BrickTalk, see how public sector teams can transform massive volumes of unstructured video and imagery into searchable, actionable insights using Databricks. We’ll walk through a model-agnostic pipeline powered by serverless GPUs and automated orchestration, enabling everything from natural language search across footage to real-time analysis and AI-generated summaries.

BrickTalks are live, virtual expert-led sessions where Databricks product leaders share real-world architectures, demos, and practical use cases for building data and AI solutions.

📅 Thursday, April 9

• 9:00 AM San Francisco

• 12:00 PM New York

• 5:00 PM London

• 9:30 PM Bengaluru

DETAILS
In this BrickTalk, see how public sector teams can transform massive volumes of unstructured video and imagery into searchable, actionable insights using Databricks. We’ll walk through a model-agnostic pipeline powered by serverless GPUs and automated orchestration—enabling everything from natural language search across footage to real-time analysis and AI-generated summaries.

Overview
Public sector agencies manage vast quantities of unstructured video and imagery from sources like traffic sensors, body cameras, and satellite feeds. The challenge lies in building scalable, cost-effective pipelines to extract actionable data from these files. This session covers an app and pipeline, built and served on Databricks, designed to automate the processing and analysis of video at scale.

Platform Architecture
Building this pipeline on Databricks centralizes governance and compute, providing three primary technical advantages:

• Serverless GPU Compute: This pipeline utilizes Databricks Serverless GPUs to handle heavy inference workloads. This eliminates the need for manual cluster configuration and ensures agencies only pay for the exact compute time required to process footage.

• Automated Orchestration via Lakeflow: The pipeline is orchestrated using Lakeflow Jobs, supporting three distinct operational modes:

  • Interactive: Triggered via a Databricks App for on-demand analysis.

  • Event-Driven: Automatically initiated by file-arrival events in cloud storage.

  • Streaming: Integrated into continuous video streams for real-time processing.

• Model Agnostic Design: The framework is decoupled from specific AI models. As computer vision evolves, agencies can swap out components (such as the detection or segmentation heads) without re-engineering the underlying data ingestion or storage layers.
  
  Core Capabilities and Use Cases

• Natural Language Search: By converting video frames into searchable metadata, users can locate specific objects or events using standard text queries (e.g., "identify all heavy machinery in the restricted zone") rather than manual review.

• Downstream Generative AI: Once segmented and indexed, data can be passed to LLMs and Agents to generate text-based summaries from hours of footage or to flag anomalies based on predefined safety or security parameters.

• Operational Efficiency: Automating the "pixels-to-metadata" workflow reduces the time-to-insight for investigators and analysts.