Aashish Pokhrel

Aashish Pokhrel आशिष

apokhrel [at] uwyo [dot] edu  ·  Laramie, WY

MS student in Computer Science at the University of Wyoming (GPA 4.0/4.0, expected May 2027), advised by Dr. Shivanand Venkanna Sheshappanavar in the Geometric Intelligence Research Lab. My research focuses on reliable multimodal intelligence, particularly language grounding, negation, and compositional reasoning in vision-language models. I am also interested in human-centered AI, efficient multimodal sensing for wearable and assistive systems, and 3D scene understanding. Before graduate school, I spent 4+ years as a data engineer building pipelines and analytics for enterprise retail and EdTech clients.


News

Publications

Research Experience

Graduate Research Assistant Aug 2025 – Present

Geometric Intelligence Research Lab, University of Wyoming  ·  Advisor: Dr. Shivanand Venkanna Sheshappanavar

  • Post-hoc negation for frozen CLIP. Developed a lightweight framework combining a text decomposer with fewer than 2M parameters and a single-parameter polarity correction. Improved multiple-choice negation accuracy from 74.3% to 79.8% on COCO and from 81.1% to 89.3% on VOC2007 against evaluated post-hoc baselines. Diagnosed limitations of rotation-based correction through bootstrap confidence intervals, per-object calibration, and angular zone analysis. Manuscript in preparation.
  • SpaceVLM-DRC. Developed a training-free retrieval approach combining LLM-based query decomposition with dynamic embedding repulsion; achieved 53.0% R@5 on MSRVTT Retrieval-Neg.
  • Cultural heritage demo. Built a web demo of negation-aware image retrieval over a large collection, running on the university HPC cluster.
  • EgoThermal. Aligned depth and thermal camera streams for an egocentric RGB-D-T dataset and benchmark for wearable assistive devices, tested the pipeline on edge hardware (NVIDIA Jetson and Raspberry Pi), and annotated about 500 images.
  • 3D benchmarks. Annotated approximately 800 samples for human evaluation of unsolvable-problem detection in 3D-LLMs (UPD-3D). Cleaned approximately 1,200 meshes in MeshLab for GroceryWild500, a novel-view-synthesis benchmark of grocery-store scenes.

Professional Experience

Senior Software Engineer I, Data Jan 2025 – Jul 2025

Ycotek AI

Clients: Torrens University  ·  Media Design School  ·  Strategic Education

  • Analyzed data flows across 12+ enterprise systems to find integration gaps, then redesigned the pipelines, improving data integration accuracy and performance by 35%.
  • Defined data-access governance requirements with stakeholders at 8 institutions and built Power BI models with row-level security, cutting client-reported data escalations by 90%.
Software Engineer, Data Jul 2021 – Dec 2024

Logic Information Systems (now acquired by Accenture)

Clients: Ralph Lauren  ·  Belk  ·  Samworth Brothers (UK)  ·  Blue Mercury  ·  Express

  • Built an automated sales-monitoring pipeline using anomaly detection, cutting analyst effort by 95% and speeding decisions by 40%.
  • Automated associate-level sales scorecards for bonus decisions, cutting reporting turnaround by 60%.

Presentations & Talks

Education

Master of Science in Computer Science

University of Wyoming  ·  GPA: 4.0/4.0

Coursework: LLMs, Deep Learning, Explainable AI, Analysis of Algorithms

Aug 2025 – May 2027 (Expected)

Bachelor of Engineering in Computer Engineering

Institute of Engineering (IoE), Tribhuvan University

Dec 2016 – Jan 2021

Honors, Awards & Funding

Skills

Languages & AI Frameworks

Python R SQL C++ PyTorch TensorFlow scikit-learn

Cloud & HPC

GCP AWS (S3, Lambda) Microsoft Azure PBS Pro / SLURM

Edge & Sensing

NVIDIA Jetson Raspberry Pi Depth & thermal camera alignment

Data & BI (prior industry)

Snowflake BigQuery PostgreSQL PySpark Airflow dbt Tableau Power BI

Tools

Git LaTeX MeshLab Scrum/Agile