Watch the stack running

Real-time object detection & tracking

Live YOLO inference detecting and tracking every person in the scene with persistent IDs: the perception core of our industrial monitoring and infrastructure-security deployments. Runs at real-time frame rates on NVIDIA Jetson-class hardware after TensorRT optimisation.

Real inference footage · Ultralytics YOLO

Vehicle tracking & traffic analytics

Multi-object tracking of vehicles in dense traffic: counting, flow analytics and incident detection for smart-city and infrastructure projects, processed entirely at the edge so raw video never leaves the site.

Real inference footage · Ultralytics YOLO

Multi-person pose estimation

Whole-body tracking with 133 keypoints per person (body, face and hands) across multiple people simultaneously, used for workplace-safety analytics, ergonomic assessment and human–robot collaboration where the robot must anticipate human movement.

Real inference footage · OpenMMLab MMPose

Pose-based fall detection

Pose analysis that recognises fall events in real time, built for workplace safety, assisted living and public-space monitoring, with privacy-preserving on-edge processing (no cloud upload).

Real inference footage · Ultralytics YOLO

Vision-guided robotic manipulation

Real robot arms executing learned, perception-driven manipulation. We develop on ROS 2 with a simulation-first methodology: behaviours are validated in Gazebo/Isaac Sim before running on physical hardware.

Real footage · Hugging Face LeRobot

Rapid robot prototyping

Low-cost robotic platforms for fast experimentation: teleoperation, data collection and policy training, the way we de-risk manipulation use-cases before committing to industrial hardware.

Real footage · Hugging Face LeRobot

Generative AI in action

Generative model output from our GenAI stack. In projects we apply the same infrastructure to sovereign RAG assistants, domain fine-tuned LLMs and synthetic-data generation for training vision models where real data is scarce.