Kanly

When personal computers became mainstream, Michael Riddle and John Walker created AutoCAD, democratizing CAD software that had previously been confined to mainframe computers. That insight, realized by software engineers with no civil engineering background, permanently changed engineering as a whole.

Today, as large language models (LLMs) become mainstream, we at Kanly are exploring what they will do for physical engineering disciplines. At its core, we are asking a single question:

How can we build safe, rigorous, and contextual agents for physical engineering disciplines?

LLMs have excelled at software engineering tasks largely by nature. They require large amounts of training data, and software engineering provides an abundance of it. Code is sequential, text-based, and publicly available, capturing information from micro-scale implementation details (syntax errors, atomic changes) to macro-scale design intent (program structure and logic). This structure is exceptionally well-suited for LLM pretraining and post-training.

There are two broad paths to creating physical engineering agents.

The first is a bottom-up approach, mirroring how software engineering agents emerged. Foundational models would be trained to understand engineering logic from the micro scale to the macro scale using large volumes of design data. In practice, this trajectory does not transfer cleanly to physical engineering disciplines. Foundational design artifacts such as CAD files, drawings, calculations, and reports, are scarce, proprietary, and poorly represented as linear token sequences. Even when such data exists, converting these artifacts into representations that support reasoning, context awareness, and features like autosuggestion is a non-trivial design problem. These constraints significantly limit the feasibility of training bottom-up physical engineering agents while preserving engineering intent and safety.

The second is a top-down approach. Here, agent capabilities begin at higher-level engineering tasks and progressively move downward in detail and autonomy. By leveraging existing data around engineering review and quality assurance processes, we can first train agents to assist QA and review engineers in evaluating designs. Over time, these agents can evolve into persistent reviewer companions for design engineers—providing continuous feedback, guidance, and eventually “click-to-accept” design edits.

Our long-term goal is a multi-agent architecture with deep technical understanding of physical engineering disciplines. By starting with engineering review agents, we simultaneously provide immediate value to our customers, collect data to train models for engineering tasks, and create a framework where future engineering design agents are driven by continuous, rigorous, and autonomous review.