Nabaraj Subedi

Papers & manuscripts

Research, in writing.

Multimodal retrieval, 3D understanding, and generative models grounded in the real world.

Google Scholar ↗ResearchGate ↗

Published & accepted

2026
  1. ConstructGPT overview: an as-built scan compared with planned BIM to detect deviations, a forecast of the next construction phase, and a compliance agent citing OSHA
    3D vision · World modelsPoster

    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 on 3D-LLM/VLA

    Paper ↗
    BibTeX
    @inproceedings{subedi2026constructgpt,
  title     = {ConstructGPT: BIM-Conditioned 3D Diffusion and World Models
               for Automated Construction Site Monitoring and Compliance Checking},
  author    = {Subedi, Nabaraj and Majeed, Y. and Abdelaty, Ahmed
               and Sheshappanavar, Shivanand Venkanna},
  booktitle = {CVPR Workshop on 3D-LLM/VLA},
  year      = {2026}
}

Preprints & under review

2026
  1. PlanSightRAG teaser: OCR and low-resolution visual retrieval fail on a WYDOT standard-plan query while patch-level ColNomic retrieval succeeds
    Multimodal RAG · Civil engineeringSecond revision

    PlanSightRAG: A Visual-First Multimodal RAG for Automating Question Answering and Compliance Checking for Civil Standard Plans

    N. Subedi, S. D. Datta, A. Abdelaty, S. V. Sheshappanavar

    Automation in Construction

    arXiv ↗Code ↗
    BibTeX
    @article{subedi2026plansightrag,
  title   = {PlanSightRAG: A Visual-First Multimodal RAG for Automating Question
             Answering and Compliance Checking for Civil Standard Plans},
  author  = {Subedi, Nabaraj and Datta, Shuvo Dip and Abdelaty, Ahmed
             and Sheshappanavar, Shivanand Venkanna},
  journal = {arXiv preprint arXiv:2608.26091},
  year    = {2026}
}
  2. t-SNE of WYDOT document chunks: categories overlap in one monolithic search space, while a domain-scoped agent searches one coherent region
    Retrieval · Agentic AIUnder review

    When More Documents Hurt RAG: Measuring and Reducing Vector Search Dilution with Domain-Scoped Retrieval

    N. Subedi, A. Abdelaty, S. V. Sheshappanavar

    arXiv preprint · 2026

    arXiv ↗Code ↗
    BibTeX
    @article{subedi2026dilution,
  title   = {When More Documents Hurt RAG: Measuring and Reducing Vector Search
             Dilution with Domain-Scoped Retrieval},
  author  = {Subedi, Nabaraj and Abdelaty, Ahmed
             and Sheshappanavar, Shivanand Venkanna},
  journal = {arXiv preprint arXiv:2606.11350},
  year    = {2026}
}
  3. Generative modeling · EEGUnder review

    MontageFlow: Physically Valid High-Fidelity EEG Generation

    S. Regmi, N. Subedi, A. Pillai, Y. Y. Wu, Y. Chen, M. V. Heinz, L. Marsch, N. C. Jacobson, A. Campbell

    arXiv preprint · 2026

    arXiv ↗
    BibTeX
    @article{regmi2026montageflow,
  title   = {MontageFlow: Physically Valid High Fidelity EEG Generation},
  author  = {Regmi, Sudarshan and Subedi, Nabaraj and Pillai, Arvind
             and Wu, Yu Yvonne and Chen, Yuliang and Heinz, Michael V.
             and Marsch, Lisa and Jacobson, Nicholas C. and Campbell, Andrew},
  journal = {arXiv preprint arXiv:2610.14507},
  year    = {2026}
}
  4. Multimodal learning · InfrastructureUnder review

    CivilAny: Multimodal Model for Civil Infrastructure Inspection and Design

    N. Subedi, A. Abdelaty, S. V. Sheshappanavar