Dec 2023 – Aug 2024 · 8 mos

• 1. Working on compression techniques like MicroFormat(MXFP and MXINT).
• 2. Pytorch 2.0 eager mode support for running LLMs on custom hardware

Nov 2021 – Nov 2023 · 2 yrs

• 1. Worked on the system design of next-generation AIC Accelerators
• 2. Working on Backend compiler optimizations like operator scheduling, memory allocation and Tiling.
• 3. Optimizing models like LLMs. Generative models and CodeGen.

Jul 2019 – Nov 2021 · 2 yrs 4 mos

• Working on the support of different DL frameworks like Tensorflow, ONNX, Caffe2 on AIC100
• Optimized kernel development on Hexagon Vector DSP for object detection algorithms
• Development of various DL operators supported by existing frameworks
• Post-training quantization techniques for achieving floating-point like accuracy.

May 2017 – Jul 2017 · 2 mos · Noida Area, India

• Objective: Extracting images from a video with higher aesthetic-value using deep learning.
• Developed a web page using Node.JS(HTML+CSS+JavaScript) for labelling of the collected images.
• Pre-processed data to convert it in required LMDB and HDF5 data format using a python script. Converted previously available SqueezeNet to regression by adding Euclidean loss, Concatenation and silence layer and finetuned it with ImageNet model using GPU enabled PyCaffe.
• The total size reduced from 35MB to 5MB, computation time from 0.4s to 0.2s with same MSE=0.02

Dec 2016 – Jan 2017 · 1 mo · Mumbai Area, India

• Objective: R&D of speech transformation techniques using signal processing
• ◦ Detected word onset and offset in a sentence using FFT energy detection technique using MATLAB
• ◦ Word vowel detection by calculating FFT energy derivative maxima using Gaussian filter derivative
• ◦ Performance evaluation of the system using software like Praat and Audacity