deepstream-sop — Claude Code skill: SOP checks on video

What deepstream-sop does

The skill documents a four-stage SOPVideoProcessor pipeline inside a GPU-accelerated FastAPI service. Stage one scores frame boundaries with a GEBD detector and cuts the stream into chunks. Stage two hands those chunks to vLLM running Cosmos Reason 1 or 2, which labels the action in each one. Stage four compares the labels against the expected sequence and reports steps that are missing or out of order.

The agent also gets the internals: Triton CAPI with the DDM model, nvds_custom_postprocess, SSE streaming, Kafka output as JSON or Protobuf. Input can be a video file, an RTSP stream or a Basler camera through Pylon, with camera emulation as an option. Everything runs in Docker next to a Kafka container, and requests hit port 8300 at /v1/chat/completions. Chunking is a per-request choice: ddm-net uses GEBD, while uniform makes fixed-length chunks and skips that stage entirely.

Both models are swappable. Point DDM_MODEL_PATH at another Triton model directory, or set VLLM_MODEL_PATH to a different model. FRAMES_PER_SIDE and SEQUENCE_BATCH control the DDM temporal window. Generic DeepStream work, object detection or tracking, NIM imports and video summarization fall outside its scope, so the agent won't pick it up for those.