Research
HarMoEny
- Token balancing system to improve time to first token and throughput by 78% over Round-Robing used by Microsoft Deepspeed Inference
- Built using PyTorch, CUDA, and NCCL as a modular system with an injector that works with any PyTorch model
- Submitted to Middleware'25 (under review)
Machine Learning Benchmark (part of MLPerf)
- Conducted research on emulating storage patterns for Google's Bidirectional Encoder Representations from Transformers (BERT) without the use of accelerators
- Achieved a 99.5% cosine similarity
- Achieved using bpftrace
Work
-
EPFL: Research Intern May 2024 - August 2024
- Supervised by Prof. Anne-Marie Kermarrec
- Conducted systematic investigation into inference bottlenecks for Mixture-of-Experts (MoE) Transformer models
- See Above
-
TraderEdge: Software Lead Dec 2022 - Present
- Improved loading speed and quality of most valuable asset by 690x and reduced churn rate by a half
- Added non-existent Jest unit tests
- Reduced downtime from hours on average to minutes, and reduced number of outages
- McGill Teaching Assistant
- Web Development Fall 2023
- Delivered 8 lectures to over 100 students, covering full-stack development technologies such as Wireshark, HTML, CSS, JavaScript, Bootstrap, XAMPP, PHP, MERN, React, SQL, and MongoDB, enabling students to build comprehensive final project websites
- Graded final projects on code quality and performance
- Operating Systems Fall 2023, Winter 2024
- Delivered 3 lectures on C and GDB to a class of over 350 students each semester
- Synthesized and graded assignments
Education
- Master of Computer Science (thesis), McGill University
- Supervised by: Prof. Oana Balmau
- Cumulative GPA: 4.0/4.0
- Relevant coursework: Distributed Systems, Advanced Computer Systems, Applied Machine Learning
- Bachelor of Computer Science, McGill University
- Cumulative GPA: 3.97/4.0
- Relevant coursework: Web Development, Advanced Algorithms, Database Systems, Software Privacy, Operating Systems
Awards
- Graduation Dean's Honours List May 2023
- NSERC URSA Summer 2023
- Faculty of Science Scholarship Jul 2021
- MES Excellence Bursary Science Apr 2021
Projects
MyDrive
- Developed a privacy preserving file storage in Rust, HTML, CSS, Javascript with small enough footprint to execute on edge devices-such as, a Raspberry Pi
- Integrated TCP/IP to allow user to communicate to their personal edge device
- Produced in-depth report of Google Drive's privacy failures and implemented solutions into MyDrive
SPH Fluid Simulator
- Researched and implemented three State-of-the-Art (SOTA) Smoothed Particle Hydrodynamics (SPH) fluid simulators using OpenGL and C++
- Based on the work of Müller, Ihmsen, Solenthaler and Koschier
- Created novel benchmark and implemented multiple discretisation strategies to improve performance
Ray Tracer Low Poly Scene
- Conceptualised and developed a full scale ray tracer from the ground up
- Implemented features such as depth of field, motion blur, multi threading, anti-aliasing, soft shadows, refractions and reflections
I did try my hand at Blender here
C Compiler with OOP Support
- Developed a full C compiler with Object-Oriented Programming (OOP) support in Java17
- Implemented all components from scratch: Lexer, Parser, Abstract Syntax Tree (AST), Code Generation, and Register Allocation via graph coloring
class person {
void greet() {
print_s((char *) "hello, there\\n");
}
}
class bob extends person {
void greet() {
print_s((char *) "hi, i'm bob");
}
}
void main() {
(new class person()).greet();
(new class bob()).greet();
}
.data
label_0_:
.asciiz "hello, there\\n"
label_1_padding:
.space 1
label_2_:
.asciiz "hi, i'm bob"
# STATIC VARIABLES
# VIRTUAL TABLES
# Virtual table for person
label_5_:
.word label_3_
# Virtual table for bob
label_6_:
.word label_4_
.data
.text
.globl main
.data
.text
label_3_:
addi $sp,$sp,-4
sw $fp,0($sp)
add $fp,$zero,$sp
addi $sp,$sp,-4
sw $ra,0($sp)
addi $sp,$fp,-4
addi $sp,$sp,-4
sw $t2,0($sp)
# Start FunCall
# Start StrLiteral
la $t2,label_0_
# End StrLiteral
# Start TypeCast
add $t2,$zero,$t2
# End TypeCast
li $v0,4
add $a0,$zero,$t2
syscall
# End FunCall
label_7_epilog:
lw $t2,0($sp)
addi $sp,$sp,4
addi $sp,$fp,4
lw $ra,-4($fp)
lw $fp,0($fp)
jr $ra
.data
.text
label_4_:
addi $sp,$sp,-4
sw $fp,0($sp)
add $fp,$zero,$sp
addi $sp,$sp,-4
sw $ra,0($sp)
addi $sp,$fp,-4
addi $sp,$sp,-4
sw $t2,0($sp)
# Start FunCall
# Start StrLiteral
la $t2,label_2_
# End StrLiteral
# Start TypeCast
add $t2,$zero,$t2
# End TypeCast
li $v0,4
add $a0,$zero,$t2
syscall
# End FunCall
label_8_epilog:
lw $t2,0($sp)
addi $sp,$sp,4
addi $sp,$fp,4
lw $ra,-4($fp)
lw $fp,0($fp)
jr $ra
.data
.text
main:
addi $sp,$sp,-4
sw $fp,0($sp)
add $fp,$zero,$sp
addi $sp,$sp,-4
sw $ra,0($sp)
addi $sp,$fp,-4
addi $a0,$zero,4
addi $v0,$zero,9
syscall
add $t3,$zero,$v0
la $t2,label_5_
sw $t2,0($t3)
lw $t2,0($t3)
lw $t2,0($t2)
addi $sp,$sp,-4
sw $t3,0($sp)
addi $sp,$sp,0
jalr $t2
addi $sp,$sp,4
addi $a0,$zero,4
addi $v0,$zero,9
syscall
add $t3,$zero,$v0
la $t2,label_6_
sw $t2,0($t3)
lw $t2,0($t3)
lw $t2,0($t2)
addi $sp,$sp,-4
sw $t3,0($sp)
addi $sp,$sp,0
jalr $t2
addi $sp,$sp,4
label_9_epilog:
addi $sp,$fp,4
lw $ra,-4($fp)
lw $fp,0($fp)
li $v0,10
syscall
