Spotrunner, a rapidly growing player in digital advertising solutions, wanted to simplify how advertisers build and launch Connected TV (CTV) campaigns. They envisioned an application capable of reading and analyzing video content, extracting IAB categories and proprietary emotion-based categories, and then using that data to plan context-rich campaigns.

To bring this vision to life, Spotrunner partnered with Tech Holding. Together, we developed an innovative context-planning agent that processes a user’s uploaded advertisement, semantically matches it to relevant content, and generates highly targeted audience segments. Beyond user uploads, the system integrates with Iris.tv to ingest up to millions of video assets and millions of hours of content—all while maintaining robust performance and accuracy through a Kubernetes environment on GKE for the main application, GCP VMs for advanced video analysis, Cloud SQL with the pgvector extension for powerful semantic search, and GCP Pub/Sub for orchestrating service integrations at scale.

Spotrunner approached Tech Holding with four core objectives:

1. Context-Aware Planning Agent
2. • Create an application that leverages multimodal Large Language Models (LLMs) to scan an uploaded advertisement.
   • Automatically categorize it according to both IAB categories and proprietary emotion-based categories.
   • Ensure these insights directly inform how campaigns are built and segmented.
2. High-Volume Video Ingestion
3. • Integrate with Iris.tv to ingest and analyze up to million videos, each requiring IAB and emotion-based categorization.
   • Maintain low latency and high throughput despite massive data volumes.
3. Scalability and Cost Efficiency
4. • Reduce the cost of video processing from $10/hour of content to a more sustainable level (ultimately hitting an estimated $0.15/hour).
   • Avoid performance bottlenecks and maintain accuracy in categorizing both IAB and emotion-based attributes.
4. Seamless User Experience
5. • Allow advertisers to preview contextually matched content.
   • Provide a chat-driven, RAG-like architecture so users can iteratively refine their campaigns before final submission to the DSP of their choice.

Without a robust, scalable infrastructure, Spotrunner risked slow processing times, escalating compute expenses, and inconsistent categorizations—potentially undermining the efficacy of its CTV campaign strategy.

Tech Holding orchestrated a multi-faceted solution to meet Spotrunner’s unique needs.

Phase One: GKE Clusters for Webhook Integration, Front-End, Analytics

We containerized Spotrunner’s core application—responsible for both user interactions and webhook integrations with Iris.tv—on Google Kubernetes Engine (GKE). This included:

• Webhooks for Iris.tv: Ensuring seamless ingestion of incoming video assets.
• Front-End & Back-End Services: Enabling advertisers to upload new content, manage campaigns, and preview matching videos.
• Analytics Platform (Superset): Deployed on these same clusters, allowing Spotrunner to monitor real-time metrics and gain insights into system performance.
• GCP Pub/Sub: Facilitating reliable, asynchronous communication among services—webhooks, video processing tasks, and other microservices—across the platform.

By leveraging GKE and Pub/Sub, we created a fault-tolerant, event-driven environment that could scale horizontally based on workload demands.

Phase Two: GCP VMs for Multimodal Video Analysis

The heavy lifting—video analysis and emotion-based segment extraction—took place on Google Cloud Platform (GCP) via VMs equipped with T4 GPUs. Our team experimented with multiple Large Language Model configurations to determine the best approach:

1. Video-to-Text & Text Generation Models
2. • Extracted descriptors from video content using custom feature extraction pipelines.
   • Performed text generation to classify each video snippet based on IAB categories and proprietary emotion metrics.
2. Embedding Models
3. • Generated high-dimensional embeddings for each video segment, capturing contextual and emotional nuances.
   • Stored these embeddings in Cloud SQL with the pgvector extension, enabling rapid, semantic lookups.
3. Custom Feature Extraction
4. • Implemented specialized modules to handle the video-specific components, ensuring the system accurately detected key emotional peaks and categories.

Phase Three: Cost Optimization & Instance Scaling

To handle massive traffic spikes—especially during ingestion of millions of assets—we employed instance scaling groups on GCP. This allowed the system to dynamically spin up or tear down VMs based on load, maintaining both throughput and cost efficiency.

Through rigorous testing, we reduced the average hourly cost of processing one hour of content from $10 to an estimated $0.15, all while maintaining high accuracy benchmarks derived from industry-leading multimodal LLMs.

Phase Four: Continuous Refinement & User Engagement

Finally, we integrated a chat-driven, RAG-like architecture into Spotrunner’s main application. As users iterated on their campaigns, the system injected contextual data gleaned from their uploaded videos—along with relevant intelligence from the Iris.tv library—directly into the conversation. This empowered advertisers to refine their strategies in real time, ensuring their final campaigns were tightly aligned with both content and audience.

Once the user confirmed all details, the system seamlessly extracted the necessary parameters and submitted the campaign to the DSP of their choice.