Abstract

Disclosed is an AI-assisted construction document authoring system in which a large language model (LLM) agent operates in conjunction with a computer-aided design (CAD) application to author drawings autonomously. The system addresses three problems that prior approaches do not adequately resolve: insufficient tool-surface depth for production-grade CAD authoring, lack of transactional integrity in multi-step authoring operations, and the inability to adapt model behavior to firm-specific drawing conventions without retraining.

The system comprises four interoperating components. First, a process-resident tool server is loaded into the CAD application's address space and exposes a plurality of named authoring endpoints to the LLM agent over an interprocess communication channel, with every state-mutating endpoint executing inside a native CAD transaction to provide atomicity, consistency, isolation, and durability. Second, a correction capture and propagation mechanism records human overrides of LLM-initiated operations as structured rules, organizes them into a two-tier hierarchy (universal rules and firm-specific rules), and injects the assembled rule block into the system prompt at session initialization — without any modification of model weights, gradient computation, or fine-tuning. Third, an autonomous visual quality control method captures the current CAD view, retrieves a reference image from a firm-private library dynamically assembled from the firm's own approved work product, and submits both to a vision-capable model for structured compliance assessment against drawing standards. Fourth, a backend service authenticates requests by firm identity and persists reference libraries, correction records, and review decisions.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS