Nabaraj Subedi

M.S. in Computer Science @ University of Wyoming (Expected May 2027)

Graduate Research Assistant working on Multimodal Retrieval-Augmented Generation, Agentic AI, and 3D World Models for built environments—including ConstructGPT.

View Projects Download Resume

Contact: nsubedi1@uwyo.edu · +1 (307) 357-6341

Nabaraj Subedi

About Me

I’m a Graduate Research Assistant at the University of Wyoming, where I build grounded, compliance-aware AI systems that combine multimodal retrieval, agentic reasoning, and 3D scene understanding. My current focus is bringing 3D world models and BIM-conditioned generation to real construction-site monitoring for transportation and civil-engineering workflows.

Education

University of Wyoming

M.S. in Computer Science · Expected May 2027

Thesis Advisor: Dr. Shivanand Venkanna Sheshappanavar (Geometric Intelligence Lab, EECS) · Co-supervisor: Dr. Ahmed Abdelaty (Civil & Arch. Eng.)

Relevant Courses: Machine Learning & Data Mining, Intro to LLMs, Intro to Deep Learning, 3D Perception for Robots

Institute of Engineering, Tribhuvan University

B.E. in Electronics, Communication & Information Engineering · 2020–2024

Full Tuition Scholarship (merit-based)

Work Experience

Graduate Research Assistant

3D World ModelsMultimodal RAGAgentic AI

University of Wyoming · Aug 2025 – Present

  • Built ConstructGPT, a BIM-conditioned 3D diffusion and construction world model with a multimodal RAG compliance agent — accepted to the CVPR 2026 3D-LLM/VLA Workshop.
  • Develop multimodal RAG and agentic AI for state-level (WYDOT) infrastructure, with evaluation pipelines that optimize retrieval accuracy and groundedness in production.

Junior Machine Learning Engineer

OpenAILangChainLangGraph

PalmMind Technology · Apr 2024 – Aug 2024

  • Built scalable RAG chatbots for EV dealers and insurance clients using orchestrator–worker routing and parallelization.
  • Improved retrieval with web scraping, OCR (Tesseract), and Redis caching for faster, context-aware responses.

Teaching Assistant

TeachingSystemsData Mining

Pashchimanchal Campus, Tribhuvan University · Sep 2024 – Jul 2025

  • Taught C programming, OOAD, Operating Systems, and Data Mining.
  • Designed labs, guided projects, and provided detailed feedback to improve practical programming skills.

Research Interests

Primary Interests

  • Multimodal Retrieval-Augmented Generation
  • Agentic AI and compliance reasoning
  • 3D World Models for the built environment
  • Vision-Language-Action & 3D foundation models

Methods & Tools

  • 3D Vision
  • Diffusion Models
  • Knowledge Graphs
  • Point Clouds
  • Vision-Language Models
  • Generative AI

Projects

ConstructGPT CVPR 2026 · Accepted

3D DiffusionWorld ModelsBIMMultimodal RAG

Research Project · University of Wyoming · 2nd 3D-LLM/VLA Workshop (Denver, CO)

  • BIM-conditioned 3D latent diffusion and construction world model with a multimodal RAG compliance agent, trained on H100 GPUs.
  • Reached 0.806 VAE IoU on Rohbau3D and 0.534 leave-one-building-out world-model F1; released a 417-sample BIM-anomaly benchmark.
  • Automates construction-site monitoring and code-compliance checking from 3D scans against BIM.

MRAG 4.0 — Cognitive Digital Twin

Multi-AgentLiDARCADCompliance

Research Project · University of Wyoming · 2025 – Present

  • Unified site video, CAD drawings, LiDAR terrain data, and operational records into a multi-agent reasoning pipeline for construction infrastructure.
  • Specialized agents for visual inspection, CAD intent analysis, and code-compliance verification—validated on WYDOT infrastructure design data.
  • Delivers reasoning capabilities unreachable with standard RAG or off-the-shelf LLM systems.

WYDOT Multimodal Knowledge-Graph Chatbot

Neo4j KGMulti-Agent RAGGCP

Research Project · University of Wyoming · Aug 2025 – Present

  • Architected a production-grade multi-agent RAG system over a Neo4j knowledge graph (150K+ nodes, 330K+ relationships) indexing 1,656+ WYDOT engineering documents with hybrid dense+sparse retrieval.
  • Deployed a serverless backend on GCP Cloud Run + Cloud SQL with an automated feedback-loop ingestion pipeline for self-optimization in production.

Citi Bike History Data Analysis

BigQueryPythonML Pipelines

Personal Project

  • Analyzed 35M+ Citi Bike trips with full data engineering: ingestion, cleaning, feature engineering, anomaly detection.
  • Built analytics/ML pipelines (Random Forest, clustering, KDE, Sankey, community detection) to extract behavior, weather impacts, and network insights.

Banking Chatbot

OpenAITool CallingEmbeddings

Company Project

  • Researched embedding strategies, cost optimization, and token-efficient prompting to improve semantic accuracy while reducing inference cost.
  • Built agentic tool-calling chatbots enabling file retrieval, SQL querying, and task execution via structured tool pipelines.

Nepali Image Captioning

TransformerInception V3Full-Stack

Undergraduate Major Project

  • Generated paragraph-length Nepali captions using a Transformer + Inception V3; trained on 20,350 translated pairs + 800 cultural heritage images.
  • Compared Transformer vs LSTM baseline; integrated into a full-stack app (React frontend, Flask/Node backend).
  • Published work in Journal of Soft Computing Paradigm (2024).

Courses & Skills

Relevant Courses

  • Machine Learning & Data Mining
  • Intro to Large Language Models
  • Intro to Deep Learning
  • 3D Perception for Robots

Infrastructure & Databases

  • Compute: NVIDIA H100/A100, GCP Cloud Run, ARCC HPC Cluster, Docker
  • Databases: Neo4j (Knowledge Graphs), Qdrant (Vector DB), PostgreSQL, BigQuery

Technical Skills

  • Languages: Python, C/C++, SQL, CUDA (basic)
  • Deep Learning: PyTorch, HuggingFace Transformers, DeepSpeed ZeRO-3, PEFT/LoRA, DiT, Diffusion Models
  • 3D Vision: Open3D, PyTorch3D, Voxel Grids, LiDAR Processing, TLS Point Clouds, NeRF (basic)
  • VLM / LLM: Qwen2.5-VL, InternVL, LLaVA, GPT-4V, Fine-tuning, RAG, ColPali, LangChain, LangGraph
  • Soft Skills: Communication, Team Collaboration, Time Management

Publications

ResearchGate
  1. ConstructGPT: BIM-Conditioned 3D Diffusion and World Models for Automated Construction Site Monitoring and Compliance Checking N. Subedi, Y. Majeed, A. Abdelaty, S. V. Sheshappanavar · CVPR 2026 Workshop: 2nd 3D-LLM/VLA Workshop — Accepted (Poster #49), Denver, CO.
  2. Nepali Image Captioning: Generating Coherent Paragraph-Length Descriptions Using Transformer Journal of Soft Computing Paradigm, 6(1), 2024. DOI
  3. Drowsiness and Crash Detection Mobile Application for Vehicle’s Safety Journal of IoT in Social, Mobile, Analytics, and Cloud, 6(1), 2024. DOI

Awards & Honors

  • Graduate Research Assistantship (GRA) — University of Wyoming · 2025 – Present
  • Full Tuition Scholarship — Institute of Engineering, Tribhuwan University · 2019 – 2024
  • High School Scholarship — SOS Hermann Gmeiner Secondary School, Gandaki · 2017 – 2019