Research in Efficient Edge AI
Independent research in efficient vision model deployment and measurement methodology, conducted through Pesgit Engineering.
Efficient Inference on Constrained Hardware
My research is in the area of efficient vision model deployment on constrained embedded hardware. Specifically, it concerns how decisions about model compression are made, and what information is actually needed to make those decisions reliably on the hardware that will run the result.
Measurement on the Device
Compression decisions for vision models are typically guided by metrics computed from a model's structure. This research examines how those metrics relate to what is actually observed when the same configurations are measured on constrained hardware under controlled conditions.
The methodology involves measuring configurations directly on target devices and building an empirical record of what different approaches actually cost to run, using that record to evaluate the reliability of the guidance used to reach each configuration.
Where This Work Stands
A paper describing this work is in preparation, alongside protection for the method.
What I Work On
- Efficient inference on constrained and embedded hardware
- Measurement methodology for compression decisions on real hardware
- Training-free model compression for vision transformers
- Reproducibility and evidence standards in applied machine learning
Research collaboration · PhD conversations · technical review · speaking
If you are working on related problems in efficient inference, model compression, or measurement methodology, or if you are exploring a research collaboration or supervision arrangement, I would be glad to hear from you.
contact@fawadsarim.net