Senior ML Research Engineer · Kempner Institute, Harvard

Yasin Mazloumi

I build the systems that train frontier models — scaling deep-learning and LLM workloads across multi-node GPU clusters, on-premises and in the cloud.

Ph.D. in Computer Science · high-performance & distributed computing
also published as Abbas Mazloumi

Portrait of Yasin Mazloumi
13Publications
70BParams · KempnerForge
Best PaperHiPC 2020
NVIDIA DLIUniv. Ambassador

// about — the short version

Turning research ambitions into systems that scale.

I'm a Senior ML Research Engineer on the Research & Engineering team at Harvard's Kempner Institute for the Study of Natural and Artificial Intelligence. I build the frameworks and tooling that let researchers train and serve deep-learning and large-language-model workloads reliably and reproducibly across the institute's GPU clusters — much of it open source. If the faculty drive the science, our team makes the machine faster, more efficient, and more reliable.

Before Harvard I taught computer science as a lecturer at UC Riverside and San Diego State, and worked on the Graph AI team at Katana Graph. I earned my Ph.D. at UC Riverside with Prof. Rajiv Gupta (RIPLE group / GRASP projects), building scalable high-performance systems for distributed graph analytics — plus master's degrees in Computer Science (UCR) and Computer Architecture (University of Tehran).

// systems — what I build with the R&E team

Open-source systems for training at scale

Much of my work at the Kempner Institute ships as open source — the frameworks researchers train on, and the tools that keep large jobs healthy on the cluster.

KempnerForge Distributed training · MIT

A PyTorch-native framework for fault-tolerant distributed training of foundation models on AI clusters. Train the same architecture from 125M to 70B parameters by swapping a config — FSDP2, tensor, expert, and pipeline parallelism, FP8, and Mixture-of-Experts, all composable. Built for long jobs on shared SLURM clusters (async checkpointing, auto-resume, NCCL health monitoring), with activation-extraction hooks for mechanistic interpretability and NeuroAI, plus vision-language training over images and video.

125M → 70B params up to 57.8% MFU · H200 FSDP2 · TP · EP · PP FP8 + MoE VLM: image + video
PyTorchFSDP2 / DTensorMixture-of-ExpertsFP8SLURMInterpretability
Multimodal research

Multimodal / VLM

Training and reproducible-experiment infrastructure for the institute's vision-language research — SLURM launch, wandb-frozen reproducibility, and shared environment caching for repeatable runs.

PyTorchVLMSLURMwandb
Details ↗
Open science · enablement

Kempner Computing Handbook

The institute's public HPC handbook — getting started on the Kempner AI cluster, SLURM, GPU computing, scaling, and performance monitoring on one of the fastest academic clusters in the world.

Jupyter BookHPCDocs
Read it ↗

// focus — what I work on

Research & engineering focus

Distributed Training & Inference

Scaling deep-learning and LLM workloads across many GPUs — throughput, parallelism, and memory.

Distributed Graph Analytics & AI

Batched and evolving iterative graph queries evaluated at cluster scale.

High-Performance & Parallel Computing

Extracting performance from heterogeneous, on-prem and cloud clusters.

GPU Micro-Architecture & Programming

CUDA, memory systems, and the hardware beneath the frameworks.

Computer Architecture

Interconnection networks and fast data delivery for many-core processors.

// publications — selected & recent

Publications & preprints

Recent preprints · under review

Selected peer-reviewed

// experience — the path

From architecture to frontier AI

  1. 2024 — Present

    Senior ML Research Engineer

    Kempner Institute · Harvard University

    Research & Engineering team — enabling large-scale training and inference on the institute's GPU clusters.

  2. 2023 — 2024

    Lecturer in Computer Science

    UC Riverside · San Diego State University

    Taught programming, operating systems, and object-oriented design.

  3. 2021 — 2022

    Software Engineer II · Graph AI

    Katana Graph

    Production distributed graph systems for large-scale analytics.

  4. 2016 — 2023

    Ph.D., Computer Science

    UC Riverside · RIPLE / GRASP (advisor: Rajiv Gupta)

    Distributed evaluation of batches of iterative graph queries. HiPC'20 Best Paper.

  5. Earlier

    M.S., Computer Architecture

    University of Tehran

    Hybrid packet/circuit-switched networks-on-chip for many-core processors.

// talks — teaching & workshops

Talks & workshops

Let's connect

Let's talk.

Grab 30 minutes on my calendar. I'm always glad to talk through an idea, a project, or just meet someone new — pick a time and you'll get a calendar invite with a meeting link.

Opens my booking page — pick any open slot.

© 2026 Yasin Mazloumi · also published as Abbas Mazloumi Design prototype · proposed Astro rebuild