At Indian Institute of Technology Mandi’s Centre for AI and Robotics, I reduced vision-language inference latency by over 20x on embedded hardware through model optimization, quantization, and custom inference engines. I also designed a tiered perception pipeline to keep the system within fixed compute and power budgets.
I built and field-tested an autonomous system that ran perception, language understanding, and control onboard without cloud dependency, and validated it across 17+ outdoor autonomous flights. On-device profiling helped me trace a gap between bench and deployed performance to sensor I/O contending with system communication.
In my smartphone hardware project, I deployed quantized vision-language models for a multimodal navigation aid. I also replaced a physical depth sensor with a monocular metric depth network feeding a real-time visual SLAM backend, with the wrapper written from scratch in C++.
