Backend Engineer, Video Analytics and Predictive

We build sealed AI appliances for factories. AJetson-class box on the plant network watches the cameras. A DGX Spark classbox with 128 GB of unified memory holds the record of what happened and readsthe tool sensors. Only text crosses between the two, and video is fetched ondemand with an audit record of every viewing. You will build the pipeline that turns camera streamsinto typed events and tool sensor streams into early alerts that name theaffected sensors. The first release is classical computer vision and timeseries. The language model layer comes after that works. WHAT YOU WILL BUILD Edge detection and tracking. RTSP from the DVRs into GStreamer or DeepStream, inference through ONNX Runtime or TensorRT,and a frame budget per camera that holds when all thirty cameras are busy. Typed events. One event per track with a stable identifier, a confidence, a one-sentence caption and a pointer to the clip, published over MQTT with a local spool that drains after a network cut. Central ingest. PostgreSQL with TimescaleDB. Deduplicate, place the event on the site plan through the camera calibration, correlate into incidents, aggregate so every dashboard panel answers under a second. Equipment Watch. Multivariate sensor streams through RFOD and CAD, so an alert carries the affected sensors and a feature attribution rather than a bare score, measured against labelled failure windows. Replay. Role check, audit record, signed ticket, WebRTC relay from the edge box to the browser, held in memory only. The design is written. The code is missing. You will be given the specification and expected to argue with it where it is wrong. We need one to three years of Python in production, video work in one of OpenCV, GStreamer or DeepStream, model inference through ONNXRuntime, TensorRT or PyTorch, SQL beyond the basics, Docker and Linux. Jetson deployment, MQTT, TimescaleDB, time series anomaly detection and time spent in a fab or plant all help. Before we talk, four hours at home: run any detector and tracker over a short video, emit one typed event per track over MQTT, write the events to PostgreSQL, and count them by class and by minute. Send the code with a note on where it breaks at thirty cameras.