Developing Knowledge-Based Engineering (KBE) applications remains a significant challenge in high-tech industries like aerospace, where front-loaded product development demands extensive automation. The manual code completion phase of these applications is particularly problematic: it consumes considerable time, requires specialized expertise in proprietary frameworks, and creates steep learning curves that limit broader adoption. While recent advances in Artificial Intelligence (AI) have revolutionized software development assistance, commercial AI systems consistently fail when working with specialized KBE frameworks, like ParaPy. In the absence of sufficient training data on proprietary frameworks, these systems produce code that appears correct but is actually non-functional.

This research validates that retrieval-augmented approaches offer a practical alternative to retraining models for specialized domains with limited data, such as ParaPy. In these approaches, AI systems dynamically access domain-specific knowledge during operation rather than relying solely on their training. This has significant implications for industries or institutions using proprietary tools where comprehensive model retraining is economically infeasible.

The research implements this approach by developing and evaluating a dual-agent framework for AI-assisted KBE application development that operates within industrial privacy and security constraints. The framework comprises a Developer Agent optimized for code generation and debugging with the ParaPy SDK, and an Educational Agent focused on ParaPy learning support and documentation. Both agents access a knowledge infrastructure that uses semantic search over indexed ParaPy documentation, curated examples, and technical references. Additionally, the Developer Agent employs verification mechanisms that progressively check code at two core levels: syntax correctness (ensuring the code follows programming language rules) and successful execution (confirming the code runs without errors).

User testing revealed how different skill levels benefit from AI assistance. Intermediate users benefited most, showing dramatic improvements in productivity and performance. Novice users achieved substantial productivity gains and reduced framework-specific (ParaPy) errors significantly, with task completion rates approaching expert baseline performance. Expert users, however, experienced slight performance degradation due to reduced code review under time pressure. The framework successfully reduced knowledge barriers for novice and intermediate users, broadening access to specialized engineering tools.

Despite these successes, the framework has persistent limitations in understanding three-dimensional spatial relationships. This is critical for KBE applications where code must define the precise position, orientation, and assembly of physical components. The framework struggles to correctly place components in space or apply proper rotational transformations. This produces code that may be syntactically correct but results in misaligned parts or incorrectly oriented features. These geometric errors require iterative refinement with human guidance, representing a fundamental limitation of the current approach and language model architectures.

While geometric reasoning limitations suggest fundamental boundaries of current AI capabilities, the demonstrated productivity improvements and reduced knowledge requirements establish a foundation for broader AI adoption in knowledge-intensive engineering domains. The framework contributes a validated operational prototype that addresses critical gaps in AI-assisted KBE development: reducing the manual coding bottleneck, lowering knowledge barriers for new users, and providing privacy-compliant deployment options. The framework functions most effectively as a development accelerator requiring expert oversight, supporting human engineers rather than replacing them.

Keywords: Knowledge-Based Engineering, Large Language Models, Code Generation, ParaPy, Retrieval-Augmented Generation, AI-Assisted Development, Aerospace Engineering, Multi-Agent Systems

It has been over 50 years since man first set foot on the Moon and proved that mankind could extend its borders beyond Earth, and yet no permanent lunar outpost has been set yet. Similar to how the construction of larger ports allowed for the exploration of the New World in the Renaissance, human exploration of space requires infrastructure beyond that required for spaceships and launchers. As our only natural satellite, the Moon is the best choice for a permanent base, one that would be capable of refuelling interplanetary missions, providing plentiful resources for the needs of people on Earth, and giving opportunities for fundamental research...