- Systems/AI engineer specializing in C++ kernels and LLM inference on NPU/GPU.
- At Microsoft (AI Frameworks, Vancouver): work on MAIA kernels and inference pipelines; build high-performance AI software in C/C++/Python; collaborate with HW/ML teams to optimize large-scale training/inference on Microsoft accelerators.
- Previously at Huawei: co-developed
sgmv/bgmvLoRA operators for vLLM (Ascend) and led database-engine performance work. - Core skills: profiling-driven SIMD, tiling, memory-layout design, throughput tuning across kernels/runtime/I/O.
- MASc, University of Waterloo (persistent memory).
- Open source: Ascend-vLLM LoRA kernels
- Tutoring: Algorithms, Data Structures, C++, Python.
Nov 2024 - Present | Microsoft — AI Frameworks | Vancouver, Canada
- Optimize C++ LLM kernels and Python bindings for Microsoft AI accelerators (MAIA); improved inference throughput/latency in internal benchmarks.
- Implemented MAIA-specific kernels, delivering ∼40% lower latency versus prior Triton implementations across targeted ops.
- Co-designed microbenchmarks, productionized changes with CI, and profiled end-to-end paths (kernels, runtime, I/O); reduced stalls via SIMD, tiling, cache-friendly layouts, and double buffering.
Sep 2023 - Nov 2024 | Huawei Technologies Canada | Burnaby, Canada
LLM Inference (May 2024 - Nov 2024):
- Co-developed Ascend NPU operators (
sgmv,bgmv) to accelerate LoRA in vLLM; validated on multi-card nodes; subsequently open-sourced by the team under the vLLM Ascend repo. - Achieved 3-5x speedups over a BMM baseline across diverse LoRA distribution scenarios.
- Built latency/throughput regression tests and CI checks; profiled kernels and runtime to remove bottlenecks.
Database Systems (Sep 2023 - May 2024):
- Led R&D for an in-memory OLAP DB; designed and implemented asynchronous scan operators.
- Increased CPU utilization by 20-40% on TPC-H queries via vectorized execution, cache-aware algorithms, and I/O overlap.
Open source: the Ascend kernels were released at
vllm-project/vllm-ascend
, documenting how the NPU-specific sgmv/bgmv paths integrate with vLLM to speed up LoRA serving on Ascend hardware.
Sep 2021 - Sep 2023 | University of Waterloo | Waterloo, Canada
Research experience on shared memory algorithms.
Designed and implemented snapshotting mechanisms for persistent memory-mapped files and detectable objects for persistent memory.
Jan 2022 - Apr 2023 | University of Waterloo | Waterloo, Canada
Helped in holding courses in Algorithm Design, Database Systems, Systems Programming, and Data Abstraction.
May 2022 - Aug 2022 | Huawei Technologies Canada Co., Ltd. | Waterloo, Canada
Worked on a collaborative project called Towards a High-velocity Permissioned Blockchain.
Oct 2018 - Jul 2021 | University of Tehran | Tehran, Iran
Assisted in more than 15 courses such as Advanced Programming and Operating Systems.
Oct 2019 - Apr 2021 | University of Tehran Science & Technology Park | Tehran, Iran
Analyzed social network behavior in the Ethereum network using graph analysis.
Jul 2019 - Sep 2019 | University of Tehran Science & Technology Park | Tehran, Iran
Participated in several Kaggle competitions and developed solutions using Scikit-Learn.
Sep 2021 - Sep 2023 | University of Waterloo | Waterloo, Canada
Grade: 94 / 100. Thesis on snapshotting mechanisms for persistent memory-mapped files. View Thesis
Sep 2016 - Jul 2021 | University of Tehran | Tehran, Iran
Grade: 90 / 100. Thesis on social network analysis within the Ethereum ecosystem.
Published in ApPLIED@PODC 2024. DOI: 10.1145/3663338.3665832
Investigates ways to enhance the reliability of persistent memory systems, focusing on snapshotting mechanisms and their role in system resilience. Introduces new snapshotting consistency models and mechanisms to improve performance and enhance system responsiveness. Provides experimental analysis demonstrating throughput and latency improvements.
Published in ApPLIED@PODC 2022. DOI: 10.1145/3524053.3542749
Explores the adaptation of multi-core algorithms to persistent memory, introducing the "Unified Detectable Sequential Specification" (UDSS), which simplifies interfaces and coding. Experiments conducted using Intel Optane memory demonstrate the performance implications of the implementation.
vLLM Ascend Kernels (sgmv, bgmv)
Co-developed NPU-specific operators to accelerate LoRA serving on Ascend hardware and integrated them into vLLM's execution path. Validated on multi-card nodes and later open-sourced by the team.
github.com/vllm-project/vllm-ascend/tree/main/csrc/kernels
- 3-5x speedups over BMM baselines for diverse LoRA distribution scenarios.
- Latency/throughput regression tests and CI checks to prevent performance regressions.
- Documented NPU-specific integration for kernel, runtime, and I/O paths.
Montage (UW) - Online Snapshotting
Research code and experiments for online snapshotting used in persistent-memory work at the University of Waterloo.
gitlab.uwaterloo.ca/mmoridi/montage-uw/-/tree/online-snapshotting
UDSS (Unified Detectable Sequential Specification)
Implementation supporting detectable objects / persistent-memory queue experiments with Prof. Golab's group.
git.uwaterloo.ca/wgolab/DSSQueue
These contributions align with my current work on C++ kernels, LLM inference, and profiling-driven optimization (SIMD, tiling, cache-friendly layouts) across kernels, runtime, and I/O.
Issued by University of Waterloo in Oct 2021
Issued by IEEE University of Tehran Student Branch in Jan 2020
Issued by IEEE University of Tehran Student Branch in Jan 2020
Issued by University of Tehran in Apr 2019
Issued by Huawei Technologies Canada in Oct 2024.
Was awarded the Best Team Award for outstanding contributions to LLM serving performance, by the President of the Canadian Research Center.
Issued by ETHGlobal in Jun 2023.
Project: Smarter Contract. View Project
Issued by Hyperlane in Jun 2023.
Project: Smarter Contract.
Issued by Waterloo Blockchain in May 2023.
Project: Glue. View Project
Issued by University of Waterloo in Sep 2021.
Machine Learning Engineer | Software Engineer | PhD
Mohammad and I crossed paths when we both arrived in Waterloo and from our first interaction, I knew he was a special individual. His warm and engaging personality immediately puts everyone at ease, creating an atmosphere of camaraderie and genuine connection.
One aspect that continuously leaves me in awe is Mohammad's passion and expertise in the field of research. His deep understanding of complex problems and ability to articulate insightful ideas truly set him apart. Our discussions on various research topics have been nothing short of intellectually stimulating, and I am constantly impressed by the depth of his knowledge.
In conclusion, I am confident that Mohammad's contributions will prove invaluable to any team fortunate enough to have him. If you need any additional information or have any questions about Mohammad's qualifications, please do not hesitate to reach out.
Web 3 Researcher | PhD Candidate in Blockchain @UWaterloo
In today's fast-paced technology landscape, it's a rare privilege to work with professionals who excel not only in their core domain but also display an unparalleled passion for learning and innovation. Mohammad is one such gem.
Mohammad really knows his stuff when it comes to Persistent Memory. He doesn't just skim the surface; he gets deep into the details and always makes sure he's following the best guidelines. He has a knack for making complicated things about Persistent Memory clear and easy to understand. On top of that, his skills in C++ are amazing. Many of us, myself included, have learned a lot from him and appreciate how he makes tough topics simpler.
Outside of these areas, Mohammad's enthusiasm for blockchain technology sets him apart. Participating in multiple hackathons, he hasn't just showcased his skills, but he's clinched victories. These achievements aren't just a testament to his individual prowess, but they also speak volumes about his ability to collaborate, innovate, and lead teams towards a common goal.
In conclusion, having the pleasure to work alongside Mohammad has been one of the most enriching experiences of my professional journey. His blend of expertise, enthusiasm, and empathy makes him an invaluable asset to any team. I genuinely look forward to witnessing the continued trajectory of his success and breakthroughs in the future. Mohammad's potential knows no bounds, and his future is luminous.
