Computer Vision and ML Systems Intern

I contributed to computer vision work on pose estimation, focusing on the part that is easy to skip and expensive to skip: knowing whether a model actually got better.

Automated evaluation pipeline

I proposed and built an automated evaluation pipeline to benchmark internal pose estimation models, using YOLO pose and TCPFormer to assess both temporal correlation and depth accuracy. Before this, comparing two models meant comparing two sets of impressions.

Introducing consistent metrics and feedback loops improved model performance by 15% on around 60% of internal systems. The pipeline was adopted by other engineering teams, which I take as the better signal: it was useful enough that people chose it over their existing habits.

Multi-camera capture system

I designed a multithreaded 12-camera system running at 30fps with low latency, used for both real-time pose estimation and data labelling. Twelve simultaneous feeds at frame rate is a synchronisation and throughput problem before it is a vision problem: dropped or skewed frames quietly corrupt any temporal metric computed downstream.

Alongside it I built labelling software with a Flask and Tkinter interface for annotation workflows, which became the foundation for later transformer-based repetition analysis.