AWS re:Invent 2025-How to build an AI Engine to generate near real time insights for videos at scale

Summary of AWS re:Invent 2025 Presentation: "How to build an AI Engine to generate near real time insights for videos at scale"

The Problem: The "Dark Data" Crisis in Video Content

• Over 80% of internet data is video, but 90% of this video content is never analyzed
• Enterprises have massive video data repositories (3-8 PB) that remain largely untapped
• Current technology does not provide a cheap and accurate way to make video content searchable and analyzable

The Solution: A Next-Generation Video Search Engine

The presenter's company is building an AI-powered video search engine with the following key capabilities:

1. Causal Event Reasoning

• Understands the causal relationships and reasons behind events in video content
• Can answer questions like "Why is that person running?" by reasoning about past events

2. Precise Temporal Grounding

• Can pinpoint the exact moments in long video recordings when specific events occur
• Allows users to find the precise 5-second clip when a contract was signed, for example

3. Omni-Modal Retrieval

• Combines and understands all modalities in video data (visual, audio, text) in a single searchable space
• Enables queries that span multiple modalities, like "Find the moment when a glass breaking sound coincided with a person running"

4. Long-Context Understanding

• Can comprehend and reason about hours, days, or even years of continuous video footage
• Enables use cases like tracking a suspect across multiple camera feeds over an extended period

Technical Approach and Infrastructure

• The company has developed a foundational video understanding model with the above capabilities
• They are using AWS services like Parallel Cluster, S3, EC2, and NVIDIA GPUs to train and deploy their models at scale
• Partnering with AWS to ensure enterprise-grade security, trust, and robustness for their solution

Real-World Applications and Use Cases

The presenter discussed several key use cases for their video search engine:

Media and Broadcasting

• Tracking and making searchable all TV broadcasts in Japan
• Enabling summarization, sentiment analysis, and event tracking for media content

Retail Intelligence

• Understanding shopper journeys and behaviors within retail spaces

Manufacturing

• Quality control, safety monitoring, and productivity tracking using video data

Security and Forensics

• Enabling law enforcement to track suspects across multiple camera feeds over time

Conclusion and Next Steps

• The company is actively working with customers across industries to fine-tune and adapt their base model
• They are pursuing security certifications and developing secure deployment options to serve enterprise customers
• Encouraging the audience to reach out and explore potential use cases for their video search technology