A CAD kernel for modern .NET built around a hybrid geometry engine that natively supports three representations:
- B-Rep — parametric surfaces (planes, conics, NURBS) wrapped in topology, for precision modeling and STEP exchange.
- Implicit — signed distance fields (SDF) composed as an AST of primitives and operators, for lattices, shells, and organic blends.
- Mesh — discrete half-edge triangle meshes, for rendering, FEA, and 3D printing.
The unified Shape API lets you model once with one vocabulary and choose the
representation at the end:
var body = Shape.Box(40, 30, 10) - Shape.Cylinder(4, 12).Translate(10, 8, 0);
BrepSolid exact = body.ToBrep(); // precision modeling, STEP export
Sdf field = body.ToImplicit(); // blends, shells, lattices
HalfEdgeMesh mesh = body.ToMesh(); // rendering, FEA, 3D printingOn top of the kernel sit a LINQ-native spatial/topology query provider, a FeatureScript-style parametric feature history, a finite-element suite (structural, thermal, modal, buckling, harmonic, transient, fatigue, topology optimisation), a library-style OpenGL viewer (desktop and WebAssembly), an MCP server, and a code-defined ECAD stack (schematic → board → routing → Gerber/Excellon fabrication, plus enclosure fit, thermal coupling, and 3D surface routing on moulded parts).
The documentation is a set of executable examples — every code snippet is compiled,
run, and rendered by the documentation build itself, so the examples cannot drift from
the code. See the docs site (built from docs/), the design rationale in
design.md, and the per-project README.md files under src/.
.NET 10 SDK. dotnet build EngrCAD.slnx, test with dotnet test EngrCAD.slnx.
EngrCAD is written by one person, with substantial help from AI coding assistants, and I would rather say so plainly than have you guess.
A hybrid geometry kernel of this scope — three interoperating engines, a full FEA suite, a renderer, and an ECAD stack — is normally the work of a team over many years, and the research behind almost any single part of it is worth multiple PhD-years on its own (surface–surface intersection, robust boolean operations, exact geometric predicates, tetrahedral meshing, the SIMP method, involute gear conjugacy, and so on). I am one person. AI assistance is what made attempting that breadth realistic at all.
The counterweight is verification, and it is deliberate. Nothing here is trusted because it "looks right": every algorithm is checked against closed-form solutions, exact identities, twin-decoder round-trips, and measured convergence orders, and those checks live in the test suite and in the executable documentation. Where the kernel cannot do something exactly, it is designed to refuse by name rather than return a plausible wrong answer. That said — read the code and the results with their origin in mind, and please report anything that looks off.
MIT © Chris Lane. Package metadata (license, URLs) is provisional at 0.1.0 and is not
yet published to nuget.org.