City Count
The Citycount project showcased an AI-based traffic counting system, emphasizing privacy-awareness and real-time evaluation directly on an edge device.
This was my inaugural project focusing on data governance and data versioning to enhance reproducibility. An internal Intake plugin was developed specifically to manage the webdataset format and other ML-related data types.
A critical aspect of the project was the ability to switch between multiple models seamlessly, necessitating the use of effective deployment tools. During this phase, I was responsible for refactoring open-source code to facilitate easy model switching. I also conducted experiments using Triton for model deployment and versioning.
The project used Ipyannotator, a framework for data annotation, to apply post-processing techniques and enhance infered data. It leveraged the Python ecosystem extensively, including PyTorch, Python Packaging, Nbdev, Pandas, Numpy, Scikit-Image, Pillow, OpenCV, Scipy, and more.