Oct 2022 – Nov 2025 · 3 yrs 1 mo · Hyderabad, Telangana, India

• My responsibilities as a ML Ops developer with OpenCL backend were looking over the smooth deployment of our drivers on the Hardware and verifying the correctness. The methodology to verify correctness was based on first in-house tests and second based on tensorflow implementation of custom models written by me. Other important responsibility includes debugging performance drops at each and every stage of deployment.

Jun 2020 – May 2024 · 3 yrs 11 mos · Hyderabad, Telangana, India

• i have learnt 3 major skills in my tenure at Qualcomm: first understanding build systems on Linux, QNX, Windows build environment. Second what are drivers and how a driver operates in offloading a job to hardware (via other layers like kernel obviously). Third and most important writing ML operation and programming its pipeline in our driver to create a layer of a ML model which can then be combined with any implementation, weather it is custom or standard (OpenCL here has an upper hand or NPU driver as a developer has freedom to create his own custom models as we provide all the basic building blocks).

May 2019 – Apr 2020 · 11 mos · Bangalore

• Kernel programing, Driver development, AI model training.
• '2' skills i attained working at Intel: Difference between Research Methodologies at a corporate vs purely for academics. My research work during masters was literature survey of papers from various academic institutions to understand and write thesis why a particular approach works and does not work. this helped to find pattern in research approaches where the probability of good results was more. this was leveraged at Intel for writing custom implementation neural network layers models which reduce training time or reduce model size.
• i learnt ml ops implementation in standard libraries of tensorflow and modified them according to various experiments based on mathematical reasoning learnt during masters courses. this helped me publish my second paper on variations of softmax.