May 2024 – Oct 2025 · 1 yr 5 mos · Bengaluru, Karnataka, India

• Leading a team focused on AI/ML compiler development. In addition to my technical contributions highlighted below, my role involves:
• i) Collaborating with cross-functional teams to gather requirements and implement new features in the compiler
• ii) Mentoring junior engineers & interns and providing technical guidance
• iii) Tech Talks and community engagement to promote Marvell's AI solutions

Jun 2022 – May 2024 · 1 yr 11 mos · Bengaluru, Karnataka, India

• Spearheaded development on an AI/ML compiler stack to optimize and deploy deep learning models for the Octeon 10 DPU.
• Some of the areas where I have majorly contributed are:
• i) Front end integration with the open source Apache TVM framework
• ii) Implemented the hardware & simulator runtime
• iii) Compiler optimizations to decrease memory footprint and improve performance for various low level ML operators (ex. convolution layer, pooling layers).
• This work enabled many important models to fit on the hardware and boosted performance by up to 10-100% across an internal test suite / benchmark
• iv) Developed a PTQ (post training quantization) framework to instrument and convert floating point models to a quantized 8-bit integer format
• v) Support for the ONNX QDQ int8 format in the compiler
• vi) Upstreamed several bug fixes to the Apache TVM project and ONNXRT
• Tech Stack - TVM, ONNX, C++, Python
• Press Release: https://www.marvell.com/company/newsroom/marvell-open-sources-octeon-10-ml-ai-accelerator-software-5g-ran.html

Jun 2017 – May 2022 · 4 yrs 11 mos · Bengaluru, Karnataka, India

• My group at MathWorks was the MATLAB Coder team.
• The product is a source-to-source compiler which compiles MATLAB code to C/C++. I also worked closely with other code generation teams that target other languages like Verilog, VHDL, CUDA, and SystemC.
• My past projects include adding support for half precision (16 bit floating point) datatype for code generation (C/C++/CUDA/LLVM) and simulation: https://www.mathworks.com/help/fixedpoint/ref/half.html
• My most recent work at MathWorks focused on translation of MATLAB code to synthesizable SystemC code : https://www.mathworks.com/help/hdlcoder/systemc-code-generation-from-matlab.html
• I developed the initial prototype and design and led the effort to deliver the first release of the MATLAB to SystemC workflow. I also worked with customers and partners to integrate their tools with the workflow.

Aug 2016 – Dec 2016 · 4 mos · Bengaluru Area, India

• Worked with the hardware verification team to develop tooling & scripts to collect testbench coverage of hardware modules in SystemVerilog.

May 2016 – Jun 2016 · 1 mo · Greater Hyderabad Area

• Worked on understanding standard cell library design for the 16nm FinFET process node.