Email  /  GitHub  /  Google Scholar  /  LinkedIn

Building On-Device AI Systems from Hardware to Software

My on-device AI work starts from the device rather than the cloud. NanoMind and Virgile combine custom wearable hardware, cameras and sensors, local inference runtimes, and multimodal models that run without relying on a network connection.

The research challenge is system integration. Hardware constraints shape model choice, memory layout, accelerator scheduling, and user-facing latency. A good edge AI system needs hardware and software to be designed together, not patched together after model selection.