Systems Engineering
AI-Native CAD Infrastructure
Building the foundation for AI systems that can reason about, generate, and manipulate engineering geometry with the precision and reliability required for production use.
01 / Thesis
Why this matters
CAD software hasn't fundamentally changed in 40 years. The same interaction patterns, the same bottlenecks, the same gap between design intent and machine execution.
Infrastructure-first
We're not building another CAD tool. We're building the infrastructure layer that makes AI-native CAD possible—then building tools on top.
Representation matters
Current CAD formats weren't designed for AI. We need representations that preserve semantic intent, support efficient reasoning, and enable deterministic replay.
Monad integration
The same cognitive architecture powering our research tools—memory, reasoning traces, verification—applied to the geometry domain.
02 / The Problem
Why current CAD fails AI
Semantic Loss
STEP files describe geometry but discard design intent. An AI can't recover why a fillet was added or what constraint drove a dimension. Every export is a one-way trip.
Feature Order Dependence
Parametric history trees create brittle dependencies. Reordering features breaks models. AI systems need representations that are robust to operation ordering.
Non-determinism
Floating-point edge cases, solver tolerances, and kernel-specific behaviors make results non-reproducible. Verification requires deterministic replay.
Closed Ecosystems
Commercial CAD APIs are proprietary, rate-limited, and expensive. Building AI systems on top requires owning the stack.
The Translation Gap
03 / EGIR
Engram Geometry Intermediate Representation
A JSON-native format designed for AI reasoning. Preserves semantic operations, enables deterministic replay, and bridges the gap between intent and geometry.
DAG Visualization
DAG-based Operations
Directed acyclic graph of CSG operations. Order-independent, verifiable, and optimizable.
Full Provenance
Every node tracks its origin—which tool, which version, which user action. Complete audit trail.
Lazy Evaluation
B-Rep computed on demand with content-addressed caching. Efficient for large assemblies and CI/CD.
04 / Architecture
Full stack overview
From user interface to geometry kernel—every layer designed for AI integration and deterministic operation.
L0: Frontend
React/Three.js
L1: AST
CSG Operations
L2: EGIR
IR + Provenance
L3: Cache
B-Rep Storage
L4: Kernel
OpenCASCADE
05 / Open Research
Problems we're solving
Active research threads where we're pushing the state of the art.
Constraint Inference
Can we infer geometric constraints from natural language and partial specifications? How do we resolve under-constrained systems?
Topology-Aware Generation
Training diffusion models on CAD that respect topological validity. Generating B-Rep directly vs. through CSG operations.
Failure Mode Clustering
Categorizing geometry kernel failures to build retry strategies. Which operations are likely to fail together?
Multi-Fidelity Reasoning
When can we reason on bounding boxes vs. mesh vs. exact B-Rep? Adaptive fidelity for different query types.
06 / Roadmap
Phased development plan
Foundation first, then intelligence, then scale. Each phase builds on the last.
Foundation
- EGIR specification finalization
- OpenCASCADE kernel integration
- TypeScript SDK scaffold
- VS Code extension prototype
Join the initiative
We're building foundational infrastructure for AI-native engineering tools. If this resonates, we'd love to hear from you.