The Structural Engineer Agent: Behind the Scenes and Planning

After co-authoring two papers on the intersection of AI and Structural Engineering—presented at the North American Tunneling Conference (NAT 2024) and the World Tunnel Congress (WTC 2025) in Sweden—I've become deeply immersed in exploring how artificial intelligence can tangibly augment our workflows as structural engineers.

Over the past four weeks, I've dedicated significant time and energy to advancing my vision for a Structural Engineer Agent. Though I've worked extensively with AI agents and developed technical software for several years, having dedicated focus on this specific domain has been immensely rewarding.

Typically, I prefer embracing "founder mode," diving straight into building. However, I've decided to openly share my thought process as part of my "build-in-public" series—both to invite feedback and to engage with a community of engineers, AI enthusiasts, and highly technical individuals interested in solving practical engineering challenges. On today’s episode, let's focus on behind-the-scenes planning.

Why a Structural Engineer Agent?

One of the core inefficiencies in structural engineering is the manual and often tedious process of extracting relevant information from standards and codes, preparing calculations, and ensuring accuracy and thoroughness. Structural engineering demands a high degree of precision, making it challenging even to fully trust spreadsheets and calculations created by other professionals—much less by an AI.

My goal with the Structural Engineer Agent is not to replace engineers but rather to augment our capabilities significantly. Imagine quickly uploading a PDF of a handwritten calculation and having it accurately transformed into a fully formatted Mathcad document. Or rapidly querying specific codes—asking, for instance, based on EM 1110-2-2104, how hydrodynamic loads are calculated for submerged structures. You could even instruct the agent to perform critical comparisons across standards, such as analyzing differences between ASME and AWWA regarding pipe thickness calculations, and generating corresponding Mathcad files.

To ensure genuine utility, the agent is built specifically to interact seamlessly with existing domain-specific tools like SAP2000, RISA, Mathcad, and CAD software, not replace them with yet another FEA model. The objective is clear: get structural engineers from 0-80% faster, offering significant productivity gains by automating routine tasks and enabling rapid customization. You’ll be able to dispatch the agent to handle tasks such as renaming variables, inserting additional checks, or refining calculations quickly and accurately.

Today's Focus: Behind-the-Scenes Planning

As engineers, our instinct is to dissect complex problems into manageable, systematic steps. In this article, I'm sharing the thought process behind building this agent. To structure this exploration, let’s discuss two main areas critical to my planning and development:

1. The Researching Paradox

Providing an AI model with the entire ACI manual isn't feasible. Not only is it inefficient and costly, but irrelevant data also introduces noise into decision-making. To tackle this, I'm employing advanced Retrieval-Augmented Generation (RAG) techniques. Initially, I started embedding without thorough data preprocessing and immediately encountered challenges—format inconsistencies, fragmented texts, mixed content (text and PNG images), and unclear formatting.

To solve these issues, I experimented extensively with preprocessing strategies. For embedding selection, I compared OpenAI and Voyage-3-large models, ultimately selecting Voyage-3-large for its impressive retrieval accuracy and superior performance on the MTEB leaderboard. Document chunking was also critical; I tested multiple chunk sizes and finally developed a systematic approach to breaking documents using bounding boxes and extracting clean, well-structured data.

Multimodal data (images and diagrams) required special handling—images were either carefully transcribed or contextually described to enable meaningful embedding and retrieval. Semantic queries were rigorously tested to ensure accurate and comprehensive retrieval, significantly enhancing the agent’s reliability and precision.

2. Domain-Specific Integration: Mathcad

A central innovation is the seamless integration with Mathcad, enabling dynamic interaction—reading and writing variables, handling complex conditional statements, and precisely formatting mathematical expressions.

A key challenge involved managing implicit units found extensively in standards like ACI and ASCE. Unlike Mathcad, which explicitly manages units internally, these standards often assume implicit units, posing difficulties in automating accurate interpretation and conversion. My solution involved developing specialized logic within the agent to detect and systematically convert these implicit units into Mathcad-friendly formats.

Another intricate challenge was handling Mathcad’s unique syntax. While defining variables programmatically was relatively straightforward, correctly managing their precise placement, page layout, and formatting within Mathcad files required a nuanced understanding of Mathcad’s underlying structures. This demanded months of iterative coding—writing and subsequently discarding more than 20,000 lines of code before finally achieving a robust and reliable solution.

What's Next?

This article marks the start of my public journey documenting the development of the Structural Engineer Agent. My intent is not to provide a simple how-to or technical documentation, but rather to share insights, successes, and lessons learned as I build openly.

In future posts, I'll delve deeper into specifics, including my unique GraphRAG approach, handling complex data retrieval scenarios, and tackling other domain-specific challenges encountered during development.

I warmly invite you to follow along, provide feedback, and—when ready—experience firsthand how the Structural Engineer Agent can redefine efficiency in our field.

The Structural Engineer Agent is currently in closed, invite-only Beta. For my friends and colleagues I've spoken with - don't worry, you get immediate access when ready :)