A full-stack finance dashboard for accounts, cards, investments, budgets, savings goals, and multi-currency reporting, with Gemini-powered insights and voice transaction entry.
Next.jsTypeScriptSupabaseGeminiRecharts
Open project senior software engineer — amd rocm
Amir Shetaia
I build the layers most software never sees — GPU drivers, deterministic solvers, and the systems underneath.
Developing GPU drivers for ML and data center workloads at AMD on the ROCm platform. Background in HPC optimization, formal methods, and verification systems.
amir@now:~$
open to interesting problems & collaborations
local timeLIVE
--:--:--
— · Eastern · UTC−5
based in
Toronto, Ontario
Canada
43.65° N
79.38° W
79.38° W
worked at · studied at
01
About
The short version of a longer story — from Mansoura to Toronto, always drawn to the layer just below the surface.
I work where correctness, performance, and hardware meet — and I like it most when the problem is hard enough that the answer has to be measured, not guessed.
Today I'm at AMD, building Linux GPU drivers for the ROCmplatform — the layer that lets machine-learning and data-center workloads actually reach the hardware. That's kernel and driver work in C/C++: shipping features for current and next-gen GPUs and chasing down the complex issues customers and QA surface.
Before AMD I made large-scale optimization solvers deterministic and fast at Huawei (C++, OpenMP, HPC), researched formal methods and LLM-assisted verification at Queen's, and spent years in embedded — automotive, robotics, and board bring-up. The constant: low-level systems where correctness and performance both have to hold.
LIVE43.65°N
Global 1st · Huawei ICT
Cloud Track — Shenzhen, 2024
0.0
MASc GPA · /4.3
0+
people reached
0
roles across systems, HPC, ML & embedded
open to interesting problems & collaborations
02
Capabilities
The tools I reach for, grouped by where they live in the stack — from silicon up to models.
C/C++PythonC#JavaRustGoSQLJavaScriptAssemblyMATLABZephyrAUTOSAR OSEmbedded LinuxMCU DebuggingBuildrootMCTPPLDMSPDMFRUCANLINI2CSPISMBus
I3CSensor ManagementAWSDockerKubernetesTerraformCI/CDDistributed SystemsMultithreadingOpenMPHPCComputer ArchitectureGDBWinDbgWiresharkOscilloscope/Logic AnalyzerTensorFlowPyTorchHugging FaceScikit-LearnYOLONLPLLMs
Languages
10C/C++PythonC#JavaRustGoSQLJavaScriptAssemblyMATLAB
Firmware and Embedded
05ZephyrAUTOSAR OSEmbedded LinuxMCU DebuggingBuildroot
Protocols and Standards
11MCTPPLDMSPDMFRUCANLINI2CSPISMBusI3CSensor Management
Cloud and DevOps
05AWSDockerKubernetesTerraformCI/CD
System Design
05Distributed SystemsMultithreadingOpenMPHPCComputer Architecture
Debug and Tools
04GDBWinDbgWiresharkOscilloscope/Logic Analyzer
AI and ML
07TensorFlowPyTorchHugging FaceScikit-LearnYOLONLPLLMs
03
Projects
Deep dives into systems, ML, and embedded work — plus the products, research tools, and experiments I've built along the way.
P-01featured
OptVerse: Deterministic Sparse Linear Solver
R&D Software Engineer · Huawei
Made Huawei's OptVerse Cholesky solver bit-for-bit reproducible across runs in C++/OpenMP without giving up HPC throughput.
bit-exact
run-to-run reproducible
C++ / OpenMP
parallel HPC
Mittelmann
benchmark-validated
problem
Parallel sparse Cholesky factorization produced run-to-run variation — non-associative floating-point reductions and non-deterministic OpenMP scheduling gave slightly different results each run, blocking reproducible benchmarking and root-cause debugging.
approach
Traced the non-determinism to parallel reduction order and floating-point accumulation order; introduced deterministic reduction/scheduling strategies with controlled accumulation, then profiled the hot paths to recover the throughput the determinism constraints cost.
result
Delivered a deterministic Cholesky solver with reproducible output across runs, validated on Hans Mittelmann's optimization benchmarks to keep OptVerse competitive on large-scale problems.
C++OpenMPHPCSparse Linear AlgebraProfiling
Hybrid log parsing system combining DeepSeek-R1:8B and Drain algorithm achieving 97.6% accuracy across 16 datasets.
0.0%
parse accuracy
0
datasets
DeepSeek-R1
+ Drain hybrid
problem
Log parsing is critical for debugging but traditional deterministic methods struggle with variable formats and new log types.
approach
Hybrid approach combining deterministic Drain parsing with LLM-driven template generation using DeepSeek-R1:8B for improved accuracy and adaptability.
result
Achieved 97.6% parsing accuracy on 16 diverse datasets, enabling more effective anomaly detection and faster root cause analysis.
LLMsDeepSeek-R1NLPPythonLog Parsing
Raspberry Pi 4 based real-time diagnostics platform combining OBD-II telemetry, MQTT messaging, YOLO object detection, and ML for driver assistance.
<0ms
alert latency
RPi 4
edge inference
OTA
update framework
problem
Vehicle diagnostics and driver assistance systems require real-time processing with edge compute constraints and reliable connectivity.
approach
Built edge ML platform on Raspberry Pi with OBD-II integration for vehicle state, MQTT for cloud connectivity, YOLO for visual perception, and TensorFlow MobileNet for efficient inference.
result
Implemented lane departure detection and collision avoidance alerts with sub-100ms latency, designed OTA update mechanism as proof-of-concept for SDV platform evolution.
Raspberry PiOBD-IIMQTTYOLOTensorFlowMobileNet
project archive
All repositories More things I've built
Products, research tools, embedded systems, and experiments beyond the featured case studies.
An Arabic-first recipe PWA and Telegram bot that turns social links, videos, photos, and text into structured recipes using Gemini and a serverless media resolver.
Next.jsSupabaseGeminiTelegramPython
Open project A fully Arabic, RTL desktop system for photography studios covering sales, customers, debts, bookings, expenses, reporting, and local-first data storage.
ElectronReactTypeScriptSQLiteTailwind CSS
Open project A self-hosted stock research system that scans the S&P 500, combines multiple signals, advises on portfolio risk, and delivers Telegram and Discord alerts without auto-trading.
PythonFastAPINext.jsPandasMarket Data
A private AI companion with streaming chat, persistent editable memories, an evolving personality, Arabic dialect controls, and multi-model Google Gemma support.
Next.jsTypeScriptGoogle GemmaAI SDKSQLite
A browser-based Tetris agent that evaluates placements in real time, improves its heuristic weights through hill climbing, and visualizes its decisions alongside a manual play mode.
ReactTypeScriptViteHeuristic SearchVitest
A layered bare-metal driver suite for ATmega32 microcontrollers, with clean MCAL and HAL interfaces spanning core peripherals and reusable hardware modules.
CAVRBare MetalMCALHAL
04
Experience
The path here: robotics clubs and embedded benches, then cloud, HPC, and research — and now GPU drivers at AMD.
Nine roles, three countries, one throughline — moving steadily toward the layers underneath: from robotics benches in Cairo to data-center GPUs in Toronto.
RoboticsEmbeddedCloudHPC · ResearchGPU Drivers
- ›Contributing to high-impact software projects supporting current and next-generation AMD GPUs.
- ›Debugging and resolving complex Linux kernel and driver issues reported by customers and QA.
- ›Designing and implementing new driver features, documenting technical decisions and trade-offs.
- ›Collaborating with compute, machine learning, and hardware teams across AMD.
- ›Engaging with the open-source community through upstream contributions and reviews.
relocated · Egypt → Canada · 2024
05
Education & Awards
Where the foundations were poured — and a few moments that told me I was on the right track.
MASc. Electrical & Computer Engineering
Queen's University2024 – 2025
Research focus: Formal Methods, Verification & Validation, Large Language Models, System Modeling, Real-Time Systems
4.3/ 4.3
BEng. Mechatronics Engineering
Mansoura University2019 – 2024
Excellence with Honours, Top 10 of class, Academic Excellence Scholarship
3.8/ 4
Awards & Honors
- HUAWEI ICT Competition (2024)2024
First Prize Global (Shenzhen, China) and Grand Prize North Africa Regional (Tunisia) in Cloud Track
- Ideal Student Award2022
Recognition for academic performance and leadership
Community
- Co-Founder2022 – 2024Mansoura Robotics Club
Organized 4 hackathons engaging 4,000+ participants in robotics and embedded systems.
- HUAWEI ICT Academy Ambassador2024 – PresentHUAWEI
Promoted ICT education and professional development in emerging markets.
06
Live Signals
Not a static résumé — live proof that I'm still building, right now.
github · contribution graph
07
References
What the people I've built and shipped with have to say.
Hoda Saleh
Volunteer Colleague, Faculty of Engineering MU
During our volunteering activities at the Faculty of Engineering, I witnessed Amir's unwavering work ethic and dedication to excellence. He's an exceptional team player and proficient communicator who readily shares expertise and collaborates toward team objectives. A true tech enthusiast with contagious enthusiasm, and I wholeheartedly recommend him for any role.
08
Off the Clock
When I'm not in the kernel, these are the worlds I get lost in — story-first, cinematic, the kind that stick with you.
PLAYER · NightBaRron1412library 8 titlesgenre story-driven AAAplatforms PC · PS · Xboxnow playing Pragmata
09
Contact
If any of this resonates — a hard problem, a role, or just comparing notes — let's talk.
Have a hard systems problem?
Drivers, performance, reproducibility, verification — or just want to compare notes. I read every message and reply within a day or two.
based in Ontario, Canada · open to interesting work