Video Analytics at 4× Throughput

01 The problem

Camera analytics were capped by inference throughput, making every additional camera an expensive hardware decision.

02 What I did

• Migrated pipelines from Triton Inference Server to NVIDIA DeepStream, retraining and converting models for TensorRT.
• Wrote C++ extensions to pull frames and detections out of DeepStream and stream them to Redis and Kafka.
• Built face detection and live livestock counting applications on the new pipeline.
• Delivered client POCs and managed large-scale video and data storage.

03 Impact

• 4× inference throughput per camera
• More streams per GPU — lower hardware cost per camera
• Real-time events delivered to downstream systems via Kafka and Redis