About

When we started working with LLMs, we kept reaching for the same pattern: give the model a command line tool. It seemed intuitive. CLI tools are composable. They process structured data. They have predictable interfaces. They produce text output that fits naturally into a context window. We were building small programs that did one thing well and piping them together, and the models seemed to work better with that than with heavier abstractions.

Nomograph Labs grew out of wanting to lean into that intuition and see where it goes. We are exploring composable tooling for AI on engineering models, building benchmarks to measure what actually helps, and trying to understand why some tool configurations work and others don't. The name is literal: a nomograph is a graphical calculation instrument that solves equations by alignment. Our tools do something similar: they make the structure of a model legible to an AI so it can answer real questions about it.

We are exploring a methodology that we think might scale across domains. Pick a formal language that engineers use to describe systems. Build a parser for it. Build composable CLI tooling on top of the parser. Build a benchmark to measure what the tooling does for AI performance. Publish the observations. Iterate.

The goal is not a single tool. It is a question: can we make the structured artifacts that engineers build (systems models, safety cases, circuit schematics, compliance documents) as legible to AI as source code is becoming? And if we can, what does that change about how complex systems get built?

SysML v2, the next-generation systems modeling language from OMG (adopted June 2025). We have built a tree-sitter grammar (192 tests, 89% external file coverage), a Rust CLI tool with MCP server built in (14 commands, 10 MCP tools, 123 tests), and a benchmark harness (132 tasks, 4 models, 14 observations).

Future directions: AADL, OSCAL, KiCad, OpenSCAD. Formal languages for physical and product system design. The source code of the physical world.