Embedded Computer Vision Engineer

Sensys Networks is expanding its intelligent transportation platform into real-time video-based detection. We are looking for a hands‑on engineer to contribute to the full detection pipeline running on embedded camera hardware --- from raw image input through object detection, tracking, and low-latency output to downstream traffic systems.

We’re building a detection system from the ground up on a modern edge AI platform. You’ll be a core contributor on a small team, working on model training, optimisation for embedded hardware, and real‑time inference. This is an opportunity to ship production detection code and grow your edge AI expertise on a system that matters.

The right person for this role is a computer vision and ML engineer first, software engineer second. You have trained detection models on real data, deployed them to edge hardware, and developed the intuition for what fails and why. You are comfortable with Python and C tooling, but the part of the job we value most is model tuning and optimisation — making models small, fast, and accurate for deployment on resource‑constrained edge devices.

• Train, fine‑tune, and iterate on detection models for our specific domain — work with datasets, augmentation, and transfer learning to get models working well on real data
• Optimise trained models for deployment on edge hardware — quantisation, pruning, and tuning to hit latency and accuracy targets on the target device
• Design and maintain the real‑time video processing pipeline from raw image input through detection, tracking, and low‑latency output
• Develop and tune object tracking to maintain consistent identity across frames and detection zones
• Work with the team on image pre‑processing decisions (exposure, white balance, stabilisation) as they affect model robustness; deep ISP tuning is handled separately
• Work closely with the Systems Engineer to integrate detection outputs into our communication infrastructure
• Contribute to camera selection and chipset evaluation for minimum viable processing requirements

• 2+ years of hands‑on computer vision or video analytics, with at least one shipped detection model on edge hardware — you’ve trained, evaluated, and iterated a model end‑to‑end
• Experience deploying models on edge‑AI hardware (Tensor
  
  RT/Jetson, Hailo, OpenVINO, Coral, Ambarella, etc.) — the workflow and constraints transfer across vendors
• Hands‑on experience training, fine‑tuning, and re‑training deep learning models for detection tasks — you are comfortable working with datasets, augmentation strategies, and transfer learning to adapt existing models to new domains or data distributions
• Proficiency with deep learning frameworks (PyTorch, Tensor Flow, ONNX) and inference optimisation tool chains (quantisation, pruning, Tensor
  
  RT, or vendor‑specific compilers)
• Solid understanding of object detection architectures (YOLO‑family, SSD, Transformers) and multi‑object tracking algorithms (SORT, DeepSORT, Byte Track)
• C comfort on embedded Linux — you can read C, have cross‑compiled, and debugged on a target board
• Strong Python for model development and tooling
• Comfortable working in an ambiguous, early‑stage product environment, communicating trade‑offs clearly and asking for direction when scope gets fuzzy

• Prior experience with traffic, automotive, or safety‑critical vision systems
• Hands‑on experience with Ambarella chipsets and their tool chains
• Experience with multi‑sensor vision systems, camera synchronisation, or coordinating detection across overlapping zones
• Experience building scene‑quality models that assess visibility degradation (weather, occlusion, lens obstruction)
• Exposure to ISP tuning, camera firmware customisation, or hardware‑aware neural architecture search
• Knowledge of H.264/H.265 encoding pipelines and their latency implications

This role offers the chance to ship production detection code on a modern platform. You’ll be a core contributor on a small team where your work directly impacts how the system performs. You’re not salvaging legacy constraints — you have real technical freedom on model selection, training strategy, and optimisation decisions. The 6‑month MVP target is aggressive and meaningful. Longer‑term, the scope expands into intersection analytics, safety metrics, and predictive modelling.

We offer competitive compensation, a collaborative environment, and the satisfaction of shipping technology that makes real‑world intersections safer and smarter.