You turn the cameras a client already has from video into decisions: what happened, where, at what time, with the clip that proves it. The video stays at the client's site.
The role
- Train and tune detection, segmentation and tracking models on real video.
- Make them run at the client's site, on cameras and recorders that already exist.
- Turn what the model sees into events with time, place and a video clip.
- Measure false positives and negatives with the client and decide which threshold can be signed off.
Your first 90 days
- Day 30
You run the current vision pipeline on test video and deliver a measured improvement.
- Day 60
You own a complete case: from the camera feed to the alert the shift lead receives.
- Day 90
A model of yours runs on a client's cameras with its error rate documented.
Stack and tools
PythonPyTorchYOLOOpenCVffmpegRTSP / ONVIF
What we look for
Must have
- You have trained or tuned a vision model and you know why it fails.
- Fluent Python and the judgment to measure, not just run examples.
- You understand a camera's limits: light, angle, resolution, compression.
- You write code that runs the same on your machine as on a client's hardware.
Nice to have
- Experience with YOLO, ByteTrack or other trackers.
- Video over RTSP or ONVIF, DVR and NVR recorders.
- Optimizing models for hardware without a big GPU.
What working here is like
What you build goes into products that operate real organizations. You work directly with the founders, with no layers between your code and production, under the same rule we give the intelligence: a person signs what is critical.