Real-Time Video Understanding AI for Safer, Smarter Digital Advertising

From Brand Risk to Real-Time Intelligence

At SingleStore Now in Seoul 2025, Dongchan Park (VP of Engineering / Co-founder) shared how PYLER is rebuilding real-time video advertising products around understanding of content. Video is now the dominant format across platforms, and advertisers face a dual mandate: protect brand trust and capture high-intent moments.

Traditional ad-tech stacks — built on batch ETL and shallow metadata — simply can’t keep up with the speed and semantic depth of modern video. Our work with SingleStore Helios reframes the problem: how do we ingest, aggregate, and deliver video understanding fast enough to match the pace of ad signals and decision-making?

PYLER’s co-founder and VP of Engineering, Dongchan Park, presented under the topic "Video Understanding AI: Real-Time Data Processing for Digital Advertising Innovation."

Two Axes of Modern Video Advertising: Brand Safety and Context

We see video ad operations evolving along two orthogonal axes. Brand Safety is the defensive line, ensuring ads never appear next to unsafe or undesirable content. And contextual targeting is the offensive engine, aligning creatives with the most relevant moments. Data from global studies and our own campaigns show that when ads run beside sensitive content, brand affinity, quality perception, and purchase intent drop dramatically. To address this, we provide domain‑specific tools for video advertisers:

• AiD (Brand Safety) :
  Prevents ads from appearing on videos that contain sensitive, inappropriate, or brand-unsafe content by analyzing the context of each segment

  
  

• AiM (Contextual Targeting) :

  Identifies moments within videos where the content meaningfully aligns with the creative’s themes and intent, and activates campaigns only in those contextually relevant, high-fit moments

Inside Video Understanding AI: From Pixels to Context and Outcomes

Video Understanding AI has to bridge two hard problems at once: raw data complexity and contextual meaning. A single video spans thousands of frames, multiple audio layers, on-screen text, captions, and metadata — all evolving over time. And effective advertising demands even more: understanding backgrounds, objects, brands, actions, and linking those semantics to real-world outcomes.

To handle this, our pipeline breaks the problem into three clear stages:

• Multi-modal Signal Extraction :
  We break each video into meaningful moments and extract all the signals that define them—visual, auditory, textual and contextual. By aligning these signals into a timeline, every second of the video becomes structured, searchable and ready for real-time understanding.

• Safety and Risk Classification (AiD) :
  We assess each moment of a video across sensitive categories and brand-specific rules, building a real-time safety layer that keeps ads away from unsuitable or high-risk contexts.
  
  

• Creative-to-moment Matching (AiM) :

  We interpret the creative itself and match it to the most fitting video moments. The result is smarter alignment between ads and content, driving more relevant, higher-performing experiences.

Building a Real-Time Ad Brain

Achieving true real-time didn’t just require a faster database — it required a fundamentally different data architecture. Our previous design, built on Aurora database and materialized views, was inherently batch-oriented: analysis results were ETL-processed, rolled up into MViews, and served through our APIs. This created unavoidable friction: 3–5 minutes of latency, 8-second query spikes under load, and operational overhead from constantly locking and refreshing MViews. By moving to Helios, we collapsed what used to be three separate systems — ingestion pipeline, OLTP store and OLAP warehouse — into a single, real-time HTAP engine. Instead of shipping data around, we now run high-velocity inserts, multi-table aggregations, and low-latency queries in one place, on fresh data.

Here’s what that unlocked:

• Unified Architecture:

  Data flows from ad platforms into batch and streaming collectors, is ingested directly into SingleStore Helios, powers our inference servers and dashboards in real time, and then feeds updated status back into the ad platforms. No more MViews, ETL refresh cycles, or warehouse hops — everything streams directly into Helios and is immediately available for analysis.

  
  

• HTAP Execution: 

  The same cluster now handles both OLTP workloads—such as frequent campaign updates and rule CRUD—and OLAP workloads, including large-scale GROUP BY, ORDER BY, and COUNT DISTINCT operations across millions of rows, all without requiring replicas or any data movement.

  
  

• Performance Transformation: 

  We observed that complex queries now run up to 6x  faster, dashboard analytics return in just 1-2 seconds, and overall resource consumption has dropped by approximately 70%.

  
  

• Product Impact: 

  Targeting rules are recalculated using inference results that are only seconds old, enabling genuinely adaptive, always-on optimization that was simply not achievable with our previous architecture.

This image illustrates how the service achieves HTAP execution and delivers a significant performance transformation.

It made real-time operational + analytical workloads converge into a single system, which is the foundation of a real-time ad brain.

What’s Next for PYLER

Going forward, we’re extending this architecture in two very practical ways:

• Trust and provenance at scale
  As AI-generated media becomes more widespread, we’re integrating emerging standards that strengthen our overall Trust & Safety posture. These improvements allow us to assess content authenticity more reliably and surface those signals to brands and platforms.

  
  

• Pushing real-time even further
  By working closely with partners like SingleStore, NVIDIA and AWS, we’re lowering end-to-end latency, increasing throughput, and preparing the system to process global video streams at massive scale.
  
  

Across all of this, our goal stays the same:
Turn the world’s video streams into something brands and users can trust — clear, real-time, and ready to act on.