Backend Engineer, Video Analytics and Predictive Maintenance

Private Advertiser; Singapore; On-site; SGD 70,000 - 110,000; Full time

Job summary

Private Advertiser seeks a software engineer to transform camera streams into typed events and sensor alerts. You will engineer edge pipelines using RTSP, GStreamer, and DeepStream, with model inference via ONNX Runtime or TensorRT.

You'll publish events over MQTT and manage a robust central ingest with PostgreSQL/TimescaleDB for real-time dashboards. The role emphasizes time-series processing, anomaly detection, and tight integration across edge and cloud components, requiring strong Python,

Qualifications

• 1-3 years of Python in production.
• Experience with video work in OpenCV, GStreamer or DeepStream.
• Model inference using ONNX Runtime, TensorRT or PyTorch.
• SQL skills beyond basics.
• Familiarity with Docker and Linux.
• Jetson deployment experience; MQTT, TimescaleDB/time-series tools.
• Exposure to time-series anomaly detection.

Responsibilities

• Build edge detection and tracking pipelines from RTSP sources using GStreamer/DeepStream.
• Perform inference with ONNX Runtime, TensorRT or PyTorch.
• Create event streams with metadata and clip pointers; publish over MQTT.
• Design central ingest with PostgreSQL/TimescaleDB for deduplication and dashboards.
• Incorporate time-series analysis and sensor data correlations for incidents.
• Implement WebRTC relay and secure audit/log records.
• Challenge specifications and iterate on requirements.

Skills

Python

OpenCV

GStreamer

DeepStream

ONNX Runtime

TensorRT

PyTorch

SQL

Docker

Linux

Jetson

MQTT

TimescaleDB

Time-series anomaly detection

Tools

PostgreSQL

Job description

About the role

You will build the pipeline that turns camera streams into typed events and tool sensor streams into early alerts that name the affected sensors. The first release is classical computer vision and time series. The language model layer comes after that works.

Key responsibilities

• 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
• Create 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
• Develop central ingest: PostgreSQL with TimescaleDB to deduplicate, place the event on the site plan through the camera calibration, correlate into incidents, aggregate so every dashboard panel answers under a second
• Implement 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
• Build Replay: role check, audit record, signed ticket, WebRTC relay from the edge box to the browser, held in memory only
• Argue with the specification where it is wrong

About you

• One to three years of Python in production
• Video work in one of OpenCV, GStreamer or DeepStream
• Model inference through ONNX Runtime, 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