About Me

I build and benchmark large language models on NVIDIA DGX A100 supercomputers. As a Postdoctoral Fellow at Southern Methodist University's O'Donnell Data Science and Research Computing Institute (ODSRCI), I work under Dr. Neena Imam (Peter O'Donnell Jr. Director of ODSRCI), which co-manages SMU's NVIDIA DGX SuperPOD, a 20-node, 160-GPU A100 AI supercomputer among the most powerful research computing platforms in the United States.

My recent work spans distributed LLM fine-tuning with multi-node DDP (1 to 12 A100 GPUs, including heterogeneous Slurm hetjob configurations), graph neural networks for innovation ecosystem analysis, and applied AI for critical infrastructure, smart grids, transportation electrification, and grid security.

Technical Focus

• Large Language Models — distributed pretraining and fine-tuning of encoder models (RoBERTa, BERT family) using PyTorch DDP, HuggingFace Transformers, and NCCL over InfiniBand. Benchmarked 1-, 2-, 4-, 8-, and 12-GPU configurations on DGX A100, including heterogeneous Slurm hetjob.

• Graph Neural Networks — large-scale graph learning for innovation ecosystem analysis across U.S. county-level patent data.

• Time-Series & Forecasting ML — short-term electric load forecasting (Kolmogorov-Arnold stacked ensembles), electricity theft detection (PFSC framework, deployed at State Grid Corporation of China), and BEV driving-range prediction (GBRP framework, IEEE T-ITS).

• HPC & Distributed Systems — NVIDIA DGX SuperPOD, Slurm scheduling, mixed-precision training (bf16/fp16), CUDA 12.x toolchains, and end-to-end data pipelines from preprocessing to checkpointed model artifacts.

Selected Engineering Work

• Distributed LLM fine-tuning on SuperPOD. Fine-tuned RoBERTa-base (125M params) with masked language modeling on the enwiki9 corpus using PyTorch DDP via Slurm. Benchmarked scaling from 1 → 12 A100 GPUs, including heterogeneous multi-node configurations (8 + 4 across two nodes via Slurm hetjob with NCCL over InfiniBand). Final eval perplexity: 3.80.

• Electricity theft detection at scale. Built end-to-end ML pipelines (preprocessing, imputation, outlier detection, ANN-based classification, evaluation with Accuracy / Precision / Recall / F1 / AUC) for 58,000 smart-meter users. The underlying PFSC framework was deployed at State Grid Corporation of China and has 10,000+ downloads on IEEE Xplore.

• DFW 2050 Mobility Planning panel dataset. Constructed a longitudinal panel covering 239 cities across 12 counties in the Dallas–Fort Worth region (2011–2024), with 22 feature groups — supporting transportation electrification and infrastructure research.

• Graph neural networks for U.S. innovation ecosystems. Designed a GNN framework that models county-level patent flows to map innovation productivity across U.S. cities. Published in IEEE Transactions on Engineering Management and IEEE Transactions on Computational Social Systems.

• BEV driving-range forecasting (GBRP framework). Gradient Boosting Range Predictor for battery electric vehicle range under varying driving conditions; manuscript under preparation for IEEE Transactions on Intelligent Transportation Systems.

Selected Publications

25 peer-reviewed papers · 500+ citations

Full lists: Google Scholar · ORCID · Lancaster Research Portal

Recent highlights:

• Data-Driven Insights: Boosting Algorithms to Uncover Electricity Theft Patterns in AMI. IEEE Transactions on Instrumentation and Measurement, 2025.
• A Novel Stacked Ensemble Framework with the Kolmogorov–Arnold Network for Short-Term Electric Load Forecasting. Elsevier Energy, 2025.
• A Predictive Analytics Framework for Policy-Driven Benchmarking and Promotion of Innovation Productivity in U.S. Cities. IEEE Transactions on Engineering Management, 2026.
• From Deserts to Hubs: A Data-Driven Framework for Mapping Innovation Productivity in the U.S.. IEEE Transactions on Computational Social Systems, 2026.

Code & Demos

• GitHub: github.com/dr-inamullahkhan — open-source code for distributed LLM fine-tuning on HPC, stacked ensemble forecasting, and graph neural networks for innovation analytics.
• Personal site: dr-inamullahkhan.github.io

Open To

I am open to research collaborations and industry roles — particularly in foundation models, distributed training, applied AI for critical infrastructure, and AI/ML for systems engineering.

For research collaborations: please use the subject line "Collaboration Proposal" and include your CV and a concise research outline.

For recruiters / industry inquiries: please use the subject line "Industry Inquiry".

Beyond Research

Based in Dallas, Texas — open to relocation. Outside work I mentor early-career researchers and led an AI/ML Bootcamp at the Texas Governor's Science & Technology Champions' Academy at SMU.

_{Left: Smart grid simulations. Middle: Machine learning model visualizations. Right: Renewable energy integration.}