The increasing capability of large language models (LLMs) creates an opportunity to redesign how users interact with computational design software. Architectural planning is a particularly relevant domain because early-stage requirements are often expressed in natural language, while the resulting design must ultimately be represented as geometry. This paper presents Omni Build Intelligence, an implemented full-stack prototype that explores an LLM-assisted workflow for generating preliminary architectural floor plans from structured user requirements. The system accepts building type, requested room count, plot width and length, and special requirements. A Flask backend constructs a domain-specific prompt and invokes a Llama 3.3 70B model through the Grok API. The response is expected in a structured textual format containing rooms, coordinates, dimensions, door relationships, and a summary.

The generated representation is parsed into easy for software to read objects and passed to a React interface. The interface provides a two-dimensional blueprint view and an isometric three-dimensional visualization. A conversion layer transforms generated room rectangles into editable wall and label primitives that can be associated with a project record. The broader platform also includes authentication, project management, plot listing and filtering, dashboard functionality, and persistence of drawing data. These components make Omni Build more than a standalone text-generation demonstration: it is a prototype for integrating generative AI into an interactive design workflow.

The research contribution of this paper is the formulation of a rigorous evaluation methodology for such a system. The proposed methodology defines an Architectural Requirement Prompt Dataset, meaningful baselines, formal geometric constraints, automated validation metrics, human evaluation criteria, ablation experiments, and statistical analysis procedures. Metrics include constraint satisfaction, boundary violations, room overlap, requirement accuracy, dimension validity, space use of space, valid-plan rate, and generation latency. A user guided study is proposed to evaluate perceived practicality and usefulness.

The study deliberately does not claim measured superiority because the repository provides implementation evidence but not a completed controlled benchmark. Instead, the paper establishes a reproducible experimental framework through which the system can be evaluated. This distinction prevents software functionality from being mistaken for research evidence. The work concludes that LLM-based architectural assistance is most appropriately positioned as an early-stage design aid whose natural-language flexibility is combined with explicit geometric validation and human review.

Keywords: large language models, generative AI, computer based architectural design, floor-plan generation, constraint based prompting, spatial reasoning, structured generation, user guided design, interactive visualization, project management