DeepStream SOP Compliance Inference
Build and operate a GPU-accelerated service that checks industrial work-step order and SOP compliance in video.
The documentation discloses flows involving files, RTSP, Basler cameras, VLM inference, Kafka, SSE, and optional RTSP output, and mentions an unprivileged default plus disconnect cleanup and required configuration. It does not systematically require user confirmation, sensitive-video handling, credential protection, rollback, or external-effect controls; Compose also exposes NGC_API_KEY, host networking, GPU access, an audio device, and SYS_PTRACE. Points are deducted for these gaps.
SKILL.md gives many key-path rules, fail-fast requirements, and cleanup instructions, while BENCHMARK.md claims test results. However, this is a static review with no committed, verifiable CI/test implementation covering the critical paths; the stack depends on GPU, DeepStream, Triton, vLLM, Pylon, and external models, and abnormal-input diagnostics are incomplete. The score therefore stays below the static ceiling.
The trigger scope, intended scenarios, and non-trigger cases for generic DeepStream, object detection, and video summarization are explicit, and input types are well described. Boundaries for Chinese-language interaction, non-GPU environments, model availability, and camera availability are not stated; core dependencies also rely on overseas model repositories, Basler downloads, and NVIDIA runtime components, with no mainland-China reachability assessment.
The skill has strong sectioning, an index, file mapping, environment-variable references, deployment, verification, and latency material, plus an owner, version, review date, and license declaration. It lacks a formal changelog, explicit maintenance/update process, fuller FAQs, and complete dependency-version governance; BENCHMARK.md records missing recommended Instructions and Examples sections, and repository license metadata is NOASSERTION.
The goal is clear and the workflow covers generation, evaluation, API use, camera requests, and latency measurement, with directly shaped examples. The source was not executed, so generated code, deployment, and inference results were not independently confirmed; GPU, commercial Pylon, and model requirements also impose substantial cost. The score is therefore conservative under static calibration.
The material provides an upstream repository reference, a DDM paper citation, detailed rules, evaluation tasks, and an internal benchmark report, giving some auditability. It lacks independently reproducible committed tests, CI evidence, and item-level third-party corroboration, so the benchmark conclusions cannot be fully verified from the supplied files alone.
- Deployment may send video or event data to external model repositories, the Basler CDN, Kafka, or RTSP; review data flows, credentials, network exposure, and authorization before production use.
- GPU, DeepStream/Triton, vLLM, model weights, and potentially commercial Pylon SDK components are required; mainland-China and no-GPU usability is unproven.
- The pass rates in BENCHMARK.md are author-supplied source material and do not replace local execution, CI, or end-to-end acceptance testing.
What does this skill do, and when should you use it?
This skill targets NVIDIA's DeepStream SOP Inference Microservice, a GPU-accelerated FastAPI service for analyzing industrial video with event boundary detection and a vision-language model. It guides agents through file, RTSP, and Basler camera inputs and identifies missing or out-of-order operating steps. The pipeline segments video with a GEBD model through Triton, classifies chunks with a vLLM-served VLM, and emits results through SSE, Kafka, and Prometheus. It fits implementation, deployment, evaluation, debugging, and chunk-latency work for this specific SOP service.
Guides configuration and operation of the FastAPI /v1/chat/completions, /v1/metrics, and /v1/metadata endpoints; builds a four-stage pipeline for DeepStream processing, Triton GEBD inference, clip segmentation, Cosmos Reason VLM classification, and SOP sequence checking; handles file, RTSP, and Basler/Pylon camera inputs; streams results over SSE and publishes JSON or NvProto messages to Kafka; and covers evaluation, verification, Docker deployment, camera emulation, and file or camera latency measurement.
- An industrial software engineer builds video-based step-order verification for an assembly line.
- A platform engineer deploys the GPU, Triton, vLLM, and Kafka service with Docker Compose.
- A computer-vision engineer debugs GEBD segmentation, VLM classification, or missing-step detection.
- A factory-integration engineer connects an RTSP or Basler camera and validates live compliance output.
- A performance engineer measures SSE time to first chunk, end-to-end latency, or camera chunk latency.
What are this skill's strengths and limitations?
- Covers implementation, deployment, evaluation, and latency measurement in one workflow.
- Clearly separates GEBD-based segmentation from fixed-length uniform chunking.
- Supports files, RTSP, and Basler/Pylon cameras, with SSE, Kafka, and Prometheus outputs.
- Documents numerous implementation constraints for cleanup, GPU-frame handling, and concurrency.
- Requires a comparatively complex NVIDIA GPU software stack including Docker, DeepStream, Triton, vLLM, and Kafka.
- Basler camera support requires the Pylon SDK; camera emulation has additional configuration requirements.
- The supplied material does not provide a complete host compatibility matrix, performance benchmark, or end-to-end prerequisite list.
- The scope is SOP sequence compliance, not generic object detection, tracking, or video summarization.
How do you install this skill?
Install the specific skill with the Skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill deepstream-sop --yes
The CLI prompts for the installation destination. The README does not specify a fixed local folder or complete host-version prerequisites.
How do you use this skill?
Load the skill directory in an Agent Skills-compatible client and trigger it with a concrete request, for example: "Build and verify a DeepStream SOP inference service that accepts RTSP input, detects missing or out-of-order assembly steps, and streams results over SSE." For live RTSP or camera inputs, the skill requires stream=true. Full implementation depends on the referenced section files and the local sop-inference-bp reference code.
How does this skill compare with similar options?
The skill defines two chunking alternatives: the default ddm-net path uses GEBD/DDM for event-boundary detection, while uniform creates fixed-length chunks and bypasses Stage 1 GEBD. It does not compare the service with other products.