Subhankar Ghosh

AI Researcher

Bellevue, Washington, United States8 yrs 5 mos experience

AI EnabledHighly Stable

Key Highlights

Ph.D. candidate specializing in Applied AI and ML.
Published research in top-tier venues.
Expertise in developing state-of-the-art AI models.

Stackforce AI infers this person is a Geospatial AI specialist with a focus on advanced machine learning techniques.

Contact

gshsubhankar@gmail.com LinkedIn

Skills

Core Skills

Generative AiAnomaly DetectionDiffusion ModelsVision TransformersGeospatial AnalysisArtificial IntelligenceSpatial Data Science

Other Skills

Agentic AIAlgorithmsApplied ResearchC++Computer VisionConditional Diffusion ModelsData MiningData StructuresDeep LearningForecastingGitGraph Neural NetworkJavaLarge Language Models (LLM)MATLAB

About

Actively seeking AI/ML/DS internship opportunities for Spring/Summer 2026. Contact Information: 📧 Email: ghosh117@umn.edu 🌐 Website: https://subhankarghosh.github.io I am a Computer Science Ph.D. candidate at the University of Minnesota, specializing in Applied AI, Machine Learning (ML), and Generative AI. With over five years of research experience, I have a strong track record of leveraging cutting-edge AI/ML techniques to solve complex interdisciplinary problems, from spatial data mining to generative modeling. My work has been published in top-tier venues such as ACM SIGSPATIAL and ACM TIST, showcasing its technical rigor and impact. My expertise spans developing state-of-the-art AI models, including Transformer-based architectures, conditional diffusion models, and probabilistic deep learning frameworks. I also have experience working with Large Language Models (LLMs), time series analysis, forecasting, anomaly detection, and computer vision, enabling me to address various challenges and applications. Proficient in tools such as Python, PyTorch, and TensorFlow, I excel in building scalable AI solutions from prototype to deployment. As a collaborative researcher, I have successfully worked with multidisciplinary teams of engineers, data scientists, and domain experts across fields like Climate Science, Economics, and GeoAI. My strengths include problem-solving, algorithm design, and effective communication, making me a valuable contributor to innovative projects.

Experience

Amazon

Applied Scientist II

Jun 2025 – Sep 2025 · 3 mos · Bellevue, Washington, United States · On-site

Developing Multi-modal Fusion and Generative AI methods for Anomaly Detection in the Amazon Science Geospatial team to improve Amazon Map data quality.

Multi-modal FusionGenerative AIAnomaly Detection

Oak ridge national laboratory

Research Scientist

Mar 2025 – May 2025 · 2 mos · Oak Ridge, Tennessee, United States · On-site

Worked with Diffusion Models & Vision Transformers for super-resolution tasks in Geospatial datasets.

Diffusion ModelsVision TransformersSuper-resolution

University of minnesota

2 roles

Graduate Research Assistant

Promoted

Jan 2022 – Mar 2025 · 3 yrs 2 mos · On-site

Developed a Generative AI framework combining Conditional Diffusion Models with Universal Kriging, achieving a 20% reduction in RMSE and a 15% improvement in spatial coherence metrics compared to baseline downscaling methods.
Analyzed historical datasets, including CMIP-6 simulation models and Copernicus satellite data, demonstrating the ability to generate high-resolution maps that align more closely with ground truth data, capturing 30% finer spatial patterns and anomalies.
Enhanced temporal consistency, achieving a 25% improvement in sequential prediction accuracy, addressing key limitations in existing models.
Delivered robust uncertainty quantification, with predictive confidence intervals offering a 40% increase in reliability for ensemble predictions of regional sea-level rise.
Developed a statistically rigorous framework for detecting significant regional and taxonomy-aware co-location patterns, reducing false positive rates by 30% using Bonferroni correction for error rate control.
Designed a taxonomy-aware mining algorithm leveraging hierarchical relationships, producing 40% more interpretable patterns compared to traditional methods.
Improved computational efficiency by cutting statistical tests and participation index computations by 50%, enabling scalable geospatial analysis.
Validated precision and robustness across synthetic and real-world datasets, identifying meaningful patterns like {Jimmy John's, McDonald's, Subway} with 20% higher recall.
Enabled actionable insights with a multi-level detection mechanism applicable to public health, retail, and urban planning use cases.
Published results in top-tier venues, showcasing superior quality, efficiency, and statistical rigor over baseline models.