Ph.D., Computer Science
University of Minnesota, Twin Cities
CV Google Scholar GitHub Hugging Face LinkedIn Twitter Email
I build physics-informed AI systems that make spatial reasoning reliable for real-world deployment: detecting GPS spoofing for safer autonomous navigation, modeling energy-efficient vehicle routing, and downscaling climate data for coastal flood-risk assessment. My research fuses deep generative models, diffusion models, transformers, physics-informed learning, and geostatistical priors to address data incompleteness, distribution shift, and violations of physical laws in spatial machine learning. I build AI systems that back each output with checkable evidence, including a real-time maritime anomaly-detection platform on AWS (Harbormaster) and a spatial reasoning agent (Spatial Atlas).
I earned my Ph.D. in Computer Science from the University of Minnesota, Twin Cities, advised by Prof. Shashi Shekhar, with committee members Prof. Vipin Kumar, Prof. Ravi Janardan, and Prof. Ying Song. My dissertation is Distortion-Aware Spatial Data Science (Doctoral Dissertation Fellowship, 2022-2023).
Actively looking for full-time opportunities.
Research Statement Teaching Statement Diversity Statement Ph.D. Thesis
Complete list, also on Google Scholar. Click a title to open the paper.
Teaching assistant at the University of Minnesota (600+ students).
Complete study collections, papers, and books I am reading. Notes include a relevance rating (out of 5); "source" links to the original.
Unofficial study notes on Hao Hoang's book about retrieval, ranking, grounding, and system design. The 60-page Notion collection has rendered equations, diagrams, full-text search, and a Notion import download. No affiliation with or endorsement by the author is claimed.
A chapter-by-chapter reading edition of Hao Hoang's guide to LLM architecture, training, inference, evaluation, and system design. Original wording, examples, code, and equations are retained. No affiliation with or endorsement by the author is claimed.
The canonical Brunton-Kutz textbook tying SVD, sparsity, ML, dynamical systems, control, ROMs, and physics-informed ML into one toolkit.
A first-principles, roofline-driven playbook for scaling Transformer training and inference across thousands of TPUs (and GPUs).
Kevin Murphy's 250-page modern RL monograph: value/policy/model-based, multi-agent, and a deep LLMs-and-RL chapter (RLHF, RLVR, PPO/GRPO/DPO).
A remote-sensing super-resolution benchmark that scores SR models by downstream task utility, not just PSNR/SSIM.
notes · notes (v1) · source
How CockroachDB delivers serializable, geo-distributed SQL transactions on commodity clouds without atomic clocks.
Theory + algorithm for scheduling LLM queries under prefix-cache reuse and TTFT limits; k-LPM cuts P99 latency vs FCFS/LPM.
LinkedIn's playbook for shrinking a 100B+ RecSys LLM 20x via distillation, structured pruning, and FP8 quant, then serving it fast.
Survey reframing code as the operational "harness" for LLM agents: reasoning, acting, state, verification, and multi-agent coordination.
Finds and fixes a GPU race condition in ZeRO++ hpZ that silently breaks 40B-70B LLM training on low-bandwidth clusters.
Google's native multimodal embedding model: one Gemini-initialized encoder maps text, image, video, and audio into a shared vector space.
1M-sample VQA benchmark probing whether vision-language models recognize dishes and origins across 30 languages and 189 countries.
Email: arunshar at umn dot edu