The right candidate's profile will clearly say things like:
- "deployed YOLO with TensorRT for real-time video"
- "built multi-camera RTSP / GStreamer pipelines"
About the Role
We run real-time people- and vehicle-analytics on live CCTV cameras. The whole system runs on-premise on GPU machines (not in the cloud). Your job is to own the video pipeline end to end:
- Pull in many live camera streams (RTSP)
- Run object detection (YOLO) optimized to run fast
- Track people and vehicles across video frames
- Turn those tracks into useful analytics (counting, zones, queues, parking)
- Keep it all running reliably, with low delay, around the clock
You will rarely train a model from scratch. You will make detection and tracking quick, stable, and production-ready.
What You'll Do
- Build and maintain real-time video pipelines (RTSP camera input → live detection → tracking → analytics)
- Optimize models for speed using TensorRT (and/or ONNX) so they run in real time on our GPUs
- Implement and tune multi-object tracking (ByteTrack / DeepSORT)
- Debug live production issues: latency,
dropped frames, GPU memory, multi-camera scaling
- Run and maintain the system on Linux GPU servers (on-premise/edge)
Required (Must Have)
- Strong Python for real production software (not just notebooks/scripts)
- Real-time video pipelines: RTSP camera ingestion, GStreamer or FFmpeg, frame handling, low-latency inference
- YOLO in production — deploying/optimizing models with TensorRT (running optimized engines, not just raw PyTorch files)
- Multi-object tracking: ByteTrack, DeepSORT, or similar
- Running computer vision on NVIDIA GPUs on-premise / edge (desktop GPUs or Jetson), on Linux
- Comfortable debugging live, always-on systems