Toronto-based ML engineer · UofT × Meta researcher
Mohammadjavad Maheronnaghsh
Machine Learning Engineer & M.A.Sc. Researcher
Efficient Deep Learning · Computer Vision · GenAI Systems
Toronto, ON · Authorized to work in Canada · PGWP-eligible
I build compact, fast, production-minded AI systems across GPU, edge, and FPGA hardware.
My recent research spans AR predistortion, lossless ML tensor compression, medical image
classification, and full-stack GenAI products.
106x
fewer parameters for Spatons AR predistortion
6x
fewer FLOPs versus CompenNet++ baseline
530 FPS
AR predistortion throughput on RTX 3090
21K+
professor profiles in Apply outreach platform
3+ yrs
across research, internships, and ML products
Featured work
Systems that connect research quality with real deployment constraints.
Efficient AR
2024 - Present
Spatons: Efficient Image Predistortion for AR Systems
10.8K-parameter patch-to-pixel network for real-time AR image predistortion, developed
at the University of Toronto in collaboration with Meta.
- Quality
- SSIM 0.944 / PSNR 27.5 dB
- Efficiency
- 106x fewer parameters, 6x fewer FLOPs
- Deployment
- RTX 3090, Arduino, Raspberry Pi 5, Zynq FPGA
PyTorch
Model Compression
Edge ML
FPGA
Apply: Agentic Academic-Outreach Platform
Full-stack GenAI web app for professor discovery, research-area matching, LLM-powered
personalized email generation, and outreach pipeline tracking.
Django
PostgreSQL
OpenRouter
Python
Football Player Tracking
Real-time multi-player detection and tracking pipeline for sports video using Segment
Anything with tracking-by-detection.
PyTorch
SAM
OpenCV
Tracking
Chromosome Classification Pipeline
Summer 2023 medical image classification work: DermaMNIST warm-up followed by a
chromosome classifier benchmarked across seven CNN backbones.
- Best results
- ResNet34, EfficientNet_v2_s, and InceptionV3 reached about 98% top-1 accuracy.
- Pipeline
- Load, resize, encode labels, split data, train, evaluate, and save models.
PyTorch
torchvision
Medical Imaging
Transfer Learning
Experience
Research depth, engineering execution, and mentoring across applied ML settings.
Sep 2024 - Present
Toronto, ON
Machine Learning Researcher — ML & XR Systems
University of Toronto · in collaboration with Meta
- Built Spatons, a 10.8K-parameter AR predistortion model matching CompenNet++ quality with 106x fewer parameters.
- Engineered an O(1)-per-pixel geometric stage and profiled real-time deployment across RTX 3090, Arduino, Raspberry Pi 5, and Zynq FPGA targets.
- Contributed to Shannonic federated-learning experiments for efficient entropy-optimal compression of ML workloads.
Jul - Sep 2023
Zista Gene Afarin
Machine Learning Intern — Computer Vision
Medical chromosome imaging
- Built a PyTorch and torchvision workflow for DermaMNIST skin-lesion classification and chromosome image classification.
- Benchmarked ResNet, DenseNet, EfficientNet, GoogleNet, InceptionV3, and AlexNet transfer-learning backbones.
- Reached about 98% top-1 accuracy with ResNet34, EfficientNet_v2_s, and InceptionV3 on the chromosome task.
Jul 2023 - Aug 2024
CISPA · Sharif
Research Intern — Fast Edge Machine Learning
CISPA Helmholtz Center, Germany · ML Lab, Sharif University of Technology
- Worked in the area of fast edge machine learning with Prof. Xiao Zhang and the Sharif ML Lab.
- Contributed to related research activity connecting efficient ML methods with resource-constrained settings.
2022 - Present
UofT · Sharif
Teaching Assistant — 20+ Course Offerings
Machine learning, LLM applications, NLP, computer vision, AI, and embedded systems
- Served as Lead/Head TA for AI and Embedded Systems courses.
- Mentored students through technical labs, assignments, and applied ML course material.
Research & publications
Efficient AI systems that learn, compress, and run under real-world constraints.
My current research focuses on efficient deep learning, computer vision, edge inference,
and hardware-aware deployment for intelligent systems.
MLSys 2026 · Accepted
SHANNONIC: Efficient Entropy-Optimal Compression for ML Workloads
Kareem Ibrahim · Mohammadjavad Maheronnaghsh · Andreas Moshovos. Contributed to federated-learning experiments for a lossless ML tensor codec with 530B state, near-Shannon-limit efficiency, and 1.3-3.1x faster communication under bandwidth constraints.
IISWC 2026 · Submitted
Spatons: Spatiotonally-Localized Building-Blocks for Efficient AR Predistortion
ECCV 2024 OOD-CV Workshop
Preprint 2025
Skills
Focused on efficient ML systems, from model design to deployment.
Languages & Frameworks
Python · C++ · CUDA · OpenMP · SQL · PyTorch · TensorFlow · HuggingFace · scikit-learn · NumPy
GenAI & LLMs
Agentic pipelines · LLM integration · Multimodal evaluation · OpenRouter · Personalized generation workflows
Efficient & Edge ML
Model compression · Efficient inference · DL acceleration · Profiling · Benchmarking · Arduino · Raspberry Pi · FPGA
Systems & MLOps
Git · Linux · AWS · REST APIs · Django · PostgreSQL · Reproducible evaluation workflows
Research Areas
Computer vision · Efficient deep learning · Reinforcement learning · NLP · Agentic AI · Robustness
Spoken Languages
English, professional · Persian/Farsi, native · Arabic, intermediate · French, beginner
Education
Academic training in ML systems, computer engineering, and applied AI.
Sep 2024 - Aug 2026 expected
GPA 3.8 / 4.0
M.A.Sc., Electrical & Computer Engineering
University of Toronto
Supervisor: Prof. Andreas Moshovos · Thesis: Efficient Image Predistortion for AR Systems in collaboration with Meta.
Nov 2020 - Jun 2024
GPA 19 / 20
B.Sc., Computer Engineering
Sharif University of Technology
Top 10% · Ranked 106 out of 150,000+ nationally.
- Eligible for the IMAE Graduate Scholarship, University of Waterloo, 2024
- Ranked 106 out of 150,000+ nationally, 2020
- National Mathematics Olympiad, Level 2, 2017-2018
