Running AI inference for large-scale camera systems often leads to massive GPU waste because each service performs redundant detection and processing on the same frames.
This talk presents a real-world architecture for building a zero-waste AI video pipeline using an event-driven design. Instead of repeating expensive inference, a single “master detector” processes frames once and publishes structured metadata through Redis Streams, allowing downstream services to reuse results for tasks such as license plate recognition, face recognition, and identity fusion.
We will explore how shared memory buffers, distributed consumers, and event streaming enable scalable multi-camera processing while dramatically reducing GPU load. The session covers architectural decisions, performance trade-offs, and lessons learned when operating AI video systems in production environments.

• Target Audience: Backend Engineers, DevOps, AI Infrastructure Engineers, System Architects.
• Core Tech Stack: Redis Streams, Python, C++, Shared Memory, GPU Inference, Docker.