Experience

Cerebras Systems

• United States
• Computer Hardware
• 200 - 300 Employee

• Tech Lead Manager + Senior Member of Technical Staff

  • Dec 2022 - Present

• Senior Member of Technical Staff

  • Aug 2022 - Dec 2022

• Member Of Technical Staff

  • Nov 2020 - Aug 2022



Interaptix Augmented Reality

• Canada
• Software Development
• 1 - 100 Employee

• Senior Software Engineer

  • Dec 2019 - Nov 2020

• Software Engineer

  • Aug 2018 - Dec 2019



Qualcomm

• United States
• Telecommunications
• 700 & Above Employee

• Machine Learning Engineer

  • Aug 2017 - Aug 2018

  Worked on Snapdragon Neural Processing Engine (SNPE), a framework for running deep neural networks on edge devices (mobile, IoT, etc.). Worked on Snapdragon Neural Processing Engine (SNPE), a framework for running deep neural networks on edge devices (mobile, IoT, etc.).



University of Toronto

• Canada
• Higher Education
• 700 & Above Employee

• Undergraduate Thesis

  • Sep 2016 - Apr 2017

  I, under supervision of Professor Roger Grosse, researched a formalized a new novel method for regularizing neural networks. The main idea behind the method was quasi-independently perturbing weights of networks. For literature review, I reviewed 25+ academic papers on various existing regularization techniques and other relevant material. By implementing the proposed framework in Theano, we achieved a test error rate of 1.35% on the MNIST dataset, which is 0.45% higher than what a standard… Show more I, under supervision of Professor Roger Grosse, researched a formalized a new novel method for regularizing neural networks. The main idea behind the method was quasi-independently perturbing weights of networks. For literature review, I reviewed 25+ academic papers on various existing regularization techniques and other relevant material. By implementing the proposed framework in Theano, we achieved a test error rate of 1.35% on the MNIST dataset, which is 0.45% higher than what a standard neural network can achieve and is comparable to the performance of Dropout. This method requires 3 orders of magnitude less memory and 15 times less training time than similar technique, DropConnect. Show less I, under supervision of Professor Roger Grosse, researched a formalized a new novel method for regularizing neural networks. The main idea behind the method was quasi-independently perturbing weights of networks. For literature review, I reviewed 25+ academic papers on various existing regularization techniques and other relevant material. By implementing the proposed framework in Theano, we achieved a test error rate of 1.35% on the MNIST dataset, which is 0.45% higher than what a standard… Show more I, under supervision of Professor Roger Grosse, researched a formalized a new novel method for regularizing neural networks. The main idea behind the method was quasi-independently perturbing weights of networks. For literature review, I reviewed 25+ academic papers on various existing regularization techniques and other relevant material. By implementing the proposed framework in Theano, we achieved a test error rate of 1.35% on the MNIST dataset, which is 0.45% higher than what a standard neural network can achieve and is comparable to the performance of Dropout. This method requires 3 orders of magnitude less memory and 15 times less training time than similar technique, DropConnect. Show less



Altera

• United States
• Semiconductors
• 700 & Above Employee

• Software Engineering Intern

  • May 2015 - Jul 2016

  • Developed a framework for automatically extracting and aggregating nightly regression-test results and archiving information in internal webpages using Perl, Python, HTML, CSS, and JavaScript • Developed an interactive graphical floorplan generator (using Python, Scalable Vector Graphics (SVG), and d3 library in JavaScript) based on an input Register Transfer Level (RTL) design • Implemented a custom Verilog preprocessor/parser in Python, using ANTLR4 (Another Tool for Language… Show more • Developed a framework for automatically extracting and aggregating nightly regression-test results and archiving information in internal webpages using Perl, Python, HTML, CSS, and JavaScript • Developed an interactive graphical floorplan generator (using Python, Scalable Vector Graphics (SVG), and d3 library in JavaScript) based on an input Register Transfer Level (RTL) design • Implemented a custom Verilog preprocessor/parser in Python, using ANTLR4 (Another Tool for Language Recognition). For this purpose, wrote a custom Verilog grammar by interpreting its Backus-Naur Form • Extended an internal tool for checking integrity of RTL designs using C++ and Verific • Wrote a script program for checking the existence of all required connections in a Verilog design • Wrote an extensive Perl program for parsing generic Excel-format (.xls or .xlsx) Test Plans provided from various groups within Altera and preparing the results in a Perl hash for querying Show less • Developed a framework for automatically extracting and aggregating nightly regression-test results and archiving information in internal webpages using Perl, Python, HTML, CSS, and JavaScript • Developed an interactive graphical floorplan generator (using Python, Scalable Vector Graphics (SVG), and d3 library in JavaScript) based on an input Register Transfer Level (RTL) design • Implemented a custom Verilog preprocessor/parser in Python, using ANTLR4 (Another Tool for Language… Show more • Developed a framework for automatically extracting and aggregating nightly regression-test results and archiving information in internal webpages using Perl, Python, HTML, CSS, and JavaScript • Developed an interactive graphical floorplan generator (using Python, Scalable Vector Graphics (SVG), and d3 library in JavaScript) based on an input Register Transfer Level (RTL) design • Implemented a custom Verilog preprocessor/parser in Python, using ANTLR4 (Another Tool for Language Recognition). For this purpose, wrote a custom Verilog grammar by interpreting its Backus-Naur Form • Extended an internal tool for checking integrity of RTL designs using C++ and Verific • Wrote a script program for checking the existence of all required connections in a Verilog design • Wrote an extensive Perl program for parsing generic Excel-format (.xls or .xlsx) Test Plans provided from various groups within Altera and preparing the results in a Perl hash for querying Show less



University of Toronto

• Canada
• Higher Education
• 700 & Above Employee

• NSERC Research Fellow

  • May 2014 - Aug 2014

  I completed a summer research fellowship under the supervision of Professor David Lie and PhD candidate James Huang. My work comprised of two main tasks. First, I developed a python project using the scrapy framework to crawl over 60,000 webpages in various vulnerability databases (such as "securityfocus.com", "securitytracker.com", and "technet.microsoft.com"). Each website's documentation followed a specific format, but my code was able to distinguish between such patterns and extract only… Show more I completed a summer research fellowship under the supervision of Professor David Lie and PhD candidate James Huang. My work comprised of two main tasks. First, I developed a python project using the scrapy framework to crawl over 60,000 webpages in various vulnerability databases (such as "securityfocus.com", "securitytracker.com", and "technet.microsoft.com"). Each website's documentation followed a specific format, but my code was able to distinguish between such patterns and extract only the desired data. This was accomplished by using XPath statements to navigate through elements and attributes of XML and HTML documents. Additionally, I utilized an internal MySQL server to save the extracted information in various tables such that they could be efficiently retrieved at a future time. As my second main task, I helped James in enhancing the Dytan Taint Analysis Framework in C++. This framework was originally written for an x86 (32-bit) architecture and we enhanced it to also work on 64-bit machines. Moreover, data tainting was only available per byte (char), however, I added the capability to taint words (for example, options in a configuration file of a software). I also implemented a mapping function from tags to memory addresses. Finally, I wrote code to instrument binary executables, using the Intel pintool. This instrumentation would allow us to bypass any tainting for some desired API methods. Show less I completed a summer research fellowship under the supervision of Professor David Lie and PhD candidate James Huang. My work comprised of two main tasks. First, I developed a python project using the scrapy framework to crawl over 60,000 webpages in various vulnerability databases (such as "securityfocus.com", "securitytracker.com", and "technet.microsoft.com"). Each website's documentation followed a specific format, but my code was able to distinguish between such patterns and extract only… Show more I completed a summer research fellowship under the supervision of Professor David Lie and PhD candidate James Huang. My work comprised of two main tasks. First, I developed a python project using the scrapy framework to crawl over 60,000 webpages in various vulnerability databases (such as "securityfocus.com", "securitytracker.com", and "technet.microsoft.com"). Each website's documentation followed a specific format, but my code was able to distinguish between such patterns and extract only the desired data. This was accomplished by using XPath statements to navigate through elements and attributes of XML and HTML documents. Additionally, I utilized an internal MySQL server to save the extracted information in various tables such that they could be efficiently retrieved at a future time. As my second main task, I helped James in enhancing the Dytan Taint Analysis Framework in C++. This framework was originally written for an x86 (32-bit) architecture and we enhanced it to also work on 64-bit machines. Moreover, data tainting was only available per byte (char), however, I added the capability to taint words (for example, options in a configuration file of a software). I also implemented a mapping function from tags to memory addresses. Finally, I wrote code to instrument binary executables, using the Intel pintool. This instrumentation would allow us to bypass any tainting for some desired API methods. Show less



University of Toronto Institute for Aerospace Studies

• Dynamic Systems Lab

• Research Assistant

  • May 2013 - Aug 2013

  My main role during this period was analyzing the behavior and controlling the position and velocity of multiple flying drones in real time by utilizing the Vicon motion capturing system. The drones in use were the commercialized Parrot AR. Drone 2.0. In order to capture their position, the drones' flying space were confined to a normal-size room, where 10 Vicon high resolution cameras were covering the entire room. By placing special markers at calculated positions on the drones, the cameras… Show more My main role during this period was analyzing the behavior and controlling the position and velocity of multiple flying drones in real time by utilizing the Vicon motion capturing system. The drones in use were the commercialized Parrot AR. Drone 2.0. In order to capture their position, the drones' flying space were confined to a normal-size room, where 10 Vicon high resolution cameras were covering the entire room. By placing special markers at calculated positions on the drones, the cameras were able to capture the position of the drones in real time. I developed code in C++ that was able to calculate the position, velocity, and aerial orientation of each drone by continuously recording the captured coordinates of each marker. Because of the high volume of incoming data, the code needed to be efficient to increase accuracy in calculations. Moreover, I wrote code that was able to command the drone to move to desired coordinates with desired velocity. Show less My main role during this period was analyzing the behavior and controlling the position and velocity of multiple flying drones in real time by utilizing the Vicon motion capturing system. The drones in use were the commercialized Parrot AR. Drone 2.0. In order to capture their position, the drones' flying space were confined to a normal-size room, where 10 Vicon high resolution cameras were covering the entire room. By placing special markers at calculated positions on the drones, the cameras… Show more My main role during this period was analyzing the behavior and controlling the position and velocity of multiple flying drones in real time by utilizing the Vicon motion capturing system. The drones in use were the commercialized Parrot AR. Drone 2.0. In order to capture their position, the drones' flying space were confined to a normal-size room, where 10 Vicon high resolution cameras were covering the entire room. By placing special markers at calculated positions on the drones, the cameras were able to capture the position of the drones in real time. I developed code in C++ that was able to calculate the position, velocity, and aerial orientation of each drone by continuously recording the captured coordinates of each marker. Because of the high volume of incoming data, the code needed to be efficient to increase accuracy in calculations. Moreover, I wrote code that was able to command the drone to move to desired coordinates with desired velocity. Show less