Write a CAD B-rep kernel in Rust and a 3,487-case conformance suite

A robotics co-founder used Claude Code to build springrep, a dependency-free Rust B-rep CAD kernel, plus a 3,487-case test suite that found hundreds of bugs before release.

Done withClaude Code· Nov 18, 2026

What they did

Tony, a co-founder of the robotics company Springcraft, kept hitting the limits of existing CAD kernels while modelling robot parts. He set out to write his own boundary-representation kernel in safe Rust with no dependencies, covering booleans, fillets, sweeps, lofts, shells, offsets, STEP import and export, and exact measurements. Alongside it he built brep-conformance, a kernel-agnostic suite of 3,487 cases whose expected answers come from closed-form maths and exact geometry rather than from another kernel's output. He credits Claude Code with making a project of this size feasible for him, and says it still took a lot of time and care.

How it went

Both the kernel and the suite are out under MIT. The suite caught hundreds of bugs and edge cases in springrep that the team's own tests had missed, and the kernel now passes all of it. The project page claims accuracy to 1e-7 mm, a validity check on every result, a WebAssembly build and the fastest median case time among the kernels tested. Tony calls it a first version and expects mistakes in the tests and adapters.

Worth knowing

The post says little about how Claude Code was used day to day; the split of work between the agent and the author is not described. The useful pattern is the independent test suite with answers derived from maths, which is what turned up the bugs. Speed and accuracy claims come from the author's own benchmark.

Try it

With Claude Code· 4 parts to fill in

Help me build [a library for a hard technical domain, e.g. a geometry kernel] in [language] with no external dependencies. Before writing the implementation, build a separate conformance suite of [number] cases whose expected answers come from closed-form maths or exact reference values, not from another implementation. Then implement features one area at a time ([list of operations]), running the full suite after each change and fixing failures before moving on. Keep a log of bugs the suite found that earlier unit tests missed. Stop and show me the results whenever a case looks wrong rather than the code, and stop when the suite passes in full.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related