AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
Senior Engineer | Compiler Optimization | High-Performance ML Systems | C++
I am a Senior Engineer specializing in compiler optimization and high-performance machine learning systems, with a focus on translating complex computational workloads into efficient, hardware-aware execution. My work centers on improving ML model performance through compiler design, optimization passes, and low-level systems engineering. I have experience across the full compiler stack, including graph transformations, auto-differentiation, and kernel-level optimizations, with an emphasis on execution latency, memory efficiency, and scalability. I regularly work with modern compiler infrastructures and ML systems such as MLIR, Triton, and LLVM-based stacks, as well as performance-critical C++ components. At Meta, I focus on optimizing ML workloads through compiler and systems-level improvements. Previously at Huawei, I contributed to end-to-end compiler stack optimization for deep learning frameworks, implementing multiple compiler passes and performance enhancements to improve execution efficiency of ML kernels. Earlier in my career, I worked on applied machine learning systems, including NLP and conversational interfaces, which gives me a strong understanding of the full ML lifecycle, from model development to production performance. I hold a PhD in Computer Science and have published in AI and intelligent systems, with additional contributions including patented work in ML kernel optimization.
Udacity
Nanodegree, Deep Reinforcement Learning
January 1, 2020 – January 1, 2020
Udacity
Nanodegree, Self-driving car
January 1, 2016 – January 1, 2017
Université du Québec en Outaouais
Doctor of Philosophy (PhD), Computer Science
January 1, 2013 – January 1, 2017
Universidad de los Andes (VE)
Master's degree, Ciencias de la computación
January 1, 2004 – January 1, 2009
Universidad Nacional Experimental del Táchira
Engineer's degree, Informática
January 1, 1997 – January 1, 2003
Meta
ML Compiler Engineer
June 1, 2022 – Present
Toronto, Ontario, Canada
Huawei Technologies
Senior Software Engineer
August 1, 2018 – June 1, 2022
Toronto, Ontario, Canada
Université du Québec en Outaouais
Chargé de cours
January 1, 2017 – May 1, 2017
Capitale-Nationale, Quebec, Canada
margoe
Co-founder - Human/Artificial Intelligence Architect
November 1, 2016 – December 1, 2018
New York City Metropolitan Area
Université du Québec en Outaouais
Assistant d'enseignement
September 1, 2016 – December 1, 2016
Capitale-Nationale, Quebec, Canada
Translate Your World International
Speech and NLP solutions
August 1, 2016 – March 1, 2018
Universite du Quebec en Outaouais
Assistant d'enseignement
January 1, 2015 – May 1, 2015
Universidad Nacional Experimental del Táchira
Docente
February 1, 2010 – October 1, 2017
Robotic Arm - Continuous control
March 1, 2020 – March 1, 2020
I implemented my own modified version of the Distributed Distributional Deep Deterministic Policy Gradient algorithm (D4PG) to find proper torque values for a robotic arm that reaches a moving target. The "brain" of the Intelligent Agents is a Neural Network with 4 Dense layers in both the Actor and the Critic. The goal of +30 average score in the Unity environment was achieved in 101 episodes.
Semantic Segmentation
September 1, 2017 – Present
In this project, I used a Fully Convolutional Network (using TensorFlow) to classify each pixel in an image. A pre-trained model on the VGG database was used as the starting point for the construction of the new network. Then, using the Kitti Road dataset (http://www.cvlibs.net/datasets/kitti/eval_road.php), I trained the network to recognize "street" pixels in an image. Skip connections, dropout layers, and regularizations approaches were used to tune the network for the best accuracy.
MPC controller
July 1, 2017 – Present
A Model Predictive Control was built using C++ to correctly drive a vehicle on the track. The MPC attempts to approximate a continuous reference trajectory by means of discrete paths between actuations. Less frequent actuations makes it harder to accurately approximate a continuous reference trajectory, too frequent could require too much computing power hence increasing response time. Therefore, part of the project involved tuning the parameters to produce a smooth driving maneuver. A polynomial of third degree must be fitted in order to compute new waypoints; this was achieved by using vectors from the Eigen library and a custom implementation of the algorithm.
Lane and vehicle detection
March 1, 2017 – Present
In this project, a video feed of a vehicle driving on a highway is given. Also, a dataset of images representing vehicles and no-vehicles was used to train a Support Vector Machine in order to classify the data from the video. Then, a python script is used to read the video feed, pass it to a pipeline in which each frame is processed with a sliding window approach to find vehicles in it. Also, another pipeline is used to detect the lane markers on the road. And finally, the detected objects are marked in the image and returned as a new video. The response time of the pipeline is very low, enough to be used and make the marks in a video stream.
Behavioral cloning
February 1, 2017 – Present
In this project, I had to record imagery data captured by a camera onboard a vehicle driven by the user in a 3D simulation environment. This data was augmented and manipulated in order to train a Deep learning network in TensorFlow. Later on, the simulation was executed in "autonomous" mode using the DL network to drive the vehicle (instead of a user). So, the project was divided into two stages: 1- Capture the data of the human driver (learn how to drive). 2- Drive the vehicle.
3D XNA environment integration with 2D Traffic simulator
June 1, 2015 – December 1, 2016
The project involves the development of an application connecting to a traffic simulator. The traffic simulator (SUMO) rules the behavior of vehicles and other entities in the simulation, it can also show in 2D the execution of it. My application connects to SUMO via a TCP/IP connection and creates a 3D representation of real-time state of the simulation. Objects such as cars, signs, traffic lights and streets are shown in 3D using the XNA graphics engine. Due to the fact that in SUMO I can't simulate collisions, one of the goals of this integration is that I can create vehicles in the 3D environment and "inject" them into the 2D simulation. This way, I'm able to design collision or other dangerous situations that involve multiple vehicles. Finally, the 3D simulator also includes a multiagent environment used by vehicles to communicate and come up with cooperative solutions for the collision scenarios.
Robot path planning
May 1, 2014 – Present
This solution is a script in Matlab that connects to a simulation environment (v-rep) in order to compute the best path to get a robot to its destination. Using a Genetic Algorithm, the program computes the path and controls the robot through it. This Khepera robot has several sensors, working with the proximity ones, their statuses are sent back to Matlab and used to know if there are any obstacles in the way. If it's the case, the program calls a Fuzzy controller to compute the best maneuver needed to get the robot away from the obstacle and back to the original path. Furthermore, in case of lost path, the program in Matlab uses again the GA to find a new path according to new location and obstacles information found on the ground by the robot.
Driver's attention detection
September 1, 2013 – April 1, 2014
Using a database of faces, both awake and sleeping (or with closed eyes), I trained a Support Vector Machine. Then, using a camera connected to a script in Matlab, I was able to obtain the stream and introduce it to previously trained SVM in order to determine if it corresponds to a sleeping or to an awake person. After several frames classified as "sleeping", the Matlab script generates an alert. The training set also included images without any person and with persons not facing front; this helped the classifier to understand the "correct" behavior of the driver as facing forward and with the open eyes. This project is part of my research in Road safety, the idea is that a vehicle can have this system on board, and a camera pointing to the driver; then, it can create audible alerts that could awake the driver in case it detects he/she is sleeping. The SVM has proven to be a great classifier with an accuracy of over 97%.
TFI
ETS
June 24, 2026 – Present
TOEFL
ETS
June 24, 2026 – Present
Duolingo French Fluency: Advanced (Estimated)
Duolingo
June 24, 2026 – Present
Natural Language Processing Specialization
Coursera
June 24, 2026 – Present
Fluidez de Duolingo en inglés: Avanzado (Estimado)
Duolingo
June 24, 2026 – Present
MCPS: Microsoft Certified Professional
Microsoft
June 24, 2026 – Present
Cultural Fit Analysis
The candidate's background is heavily focused on Machine Learning, AI, and backend/compiler optimization. While this demonstrates strong technical depth, the target role is 'Frontend Developer'. There is a significant mismatch between the candidate's demonstrated experience and the target role. The projects and professional experience are almost exclusively in AI/ML, robotics, and compiler engineering, with no explicit mention of frontend technologies, frameworks (e.g., React, Angular, Vue), or UI/UX design principles. This indicates a low cultural fit for a dedicated frontend role, as their expertise lies in a completely different domain.
Soft Skills & Operational Fit
The candidate's extensive teaching and mentorship experience suggests strong communication and leadership skills. Their co-founder role indicates entrepreneurial spirit and problem-solving abilities. The research background implies a methodical and analytical approach to challenges. However, the provided data does not offer direct insights into stress handling or team collaboration beyond the general description of roles.