Nabaraj Subedi’s Website
I am an M.S. student in Computer Science at the University of Wyoming, advised by Dr. Shivanand Venkanna Sheshappanavar (Geometric Intelligence Lab, EECS) and co-supervised by Dr. Ahmed Abdelaty (Civil & Architectural Engineering). My work is in multimodal retrieval-augmented generation, agentic AI, and 3D world models, with a focus on building grounded, compliance-aware systems for the built environment.
Research interests
- Multimodal retrieval-augmented generation
- Agentic AI and compliance reasoning
- 3D world models and generative 3D
- Vision–language models
Current and recent projects
ConstructGPT. A BIM-conditioned 3D latent diffusion and construction world model paired with a multimodal RAG compliance agent, trained on H100 GPUs. It reaches 0.806 VAE IoU on Rohbau3D and 0.534 leave-one-building-out world-model F1, and comes with a 417-sample BIM-anomaly benchmark. The goal is to automate construction-site monitoring and code-compliance checking directly from 3D scans against BIM. Accepted to the 2nd 3D-LLM/VLA Workshop at CVPR 2026.
Visual-first retrieval for civil standard plans. I work on retrieval that reads engineering drawings as images rather than flattening them to text, so question answering and compliance checking stay grounded in what the plan actually shows. A related line of work studies why adding more documents can hurt RAG, and uses domain-scoped, model-agnostic retrieval to keep vector search from diluting the signal.
WYDOT knowledge-graph assistant. A multi-agent RAG system over a Neo4j knowledge graph (150K+ nodes, 330K+ relationships) indexing 1,656+ Wyoming Department of Transportation engineering documents, with hybrid dense and sparse retrieval and a serverless backend on Google Cloud Run.
Selected publications
- PlanSightRAG: A Visual-First Multimodal RAG for Automating Question Answering and Compliance Checking for Civil Standard Plans arXiv preprint arXiv:2608.26091, 2026
- When More Documents Hurt RAG: Mitigating Vector Search Dilution with Domain-Scoped, Model-Agnostic Retrieval arXiv preprint arXiv:2606.11350, 2026
- ConstructGPT: BIM-Conditioned 3D Diffusion and World Models for Automated Construction Site Monitoring and Compliance Checking 2nd 3D-LLM/VLA Workshop, CVPR 2026 — accepted, Denver, CO
Collaboration
I am open to collaborations in multimodal AI, 3D vision, and applications to civil and transportation infrastructure — especially projects that connect rigorous methods with real engineering practice. Email: nsubedi1@uwyo.edu