Spatial Lab / Spatial Intelligence

Spatial Lab is an AI-native CAD and boundary-representation research platform for inspecting, understanding, modifying, and validating exact 3D geometry. Boundary representation, or BRep, describes a solid through its surfaces, faces, edges, and their connections. That structure matters: an image can suggest what an object looks like, but it cannot establish whether a hole, joint, or attachment survives an edit.

The project connects this geometric layer to feature relationships, constraints, and edit history. Its larger aim is to help AI reason about what parts mean, what they do together, and what may safely change within an interactive CAD workflow.

Most generative systems describe form from the outside. Spatial Lab asks whether a model can reason about the exact topology and constraints that make geometry useful. A lamp is not just a silhouette: it is a base, links, pins, a head, and relationships that allow it to move.

The central distinction is between a learned interpretation and a geometric fact. AI can propose a mechanism or an edit; deterministic tools must check the geometry. The system should make that distinction visible rather than hide uncertainty behind a convincing answer.

I developed the product direction, system architecture, geometry research, interface design, and evaluation methods. The work connects model inspection, bounded editing, and verification into a workflow a designer can review.

The interface gives the original and candidate geometry a shared frame of reference, exposes component selection and joint controls, and keeps verification evidence beside the proposed change. The development record ties capabilities to explicit tasks, review gates, and evidence rather than treating a finished-looking interface as proof that the underlying problem is solved.

The workflow starts with exact BRep geometry; the display mesh is only a view of that model, not the source of truth. Faces, edges, topology, and features supply structured information for reasoning about parts and their relationships. The Teaching Geometry Meaning diagram maps the broader research direction: geometric structure becomes a graph, geometry patterns support object concepts, and those concepts support reasoning.

An agent proposes a bounded operation. Deterministic geometry tools execute and check the candidate before it is accepted. In the lamp study, the proposed edit shortens the upper arm from 85 mm to 55 mm while preserving the head attachment and both joint interfaces. The candidate view overlays the original position so the change can be inspected directly.

The four new views tell one continuous story. The first establishes the six-solid object and its verification evidence. The second explores the articulated pose through base and elbow joint controls. The third shows the shorter-arm candidate against the original geometry. The fourth selects the editable upper arm and exposes its provenance.

The prototype reports 11 of 11 acceptance gates passed, four checked interfaces, two revolute joints, 0.30 mm pin/bore clearance, and 0.00 mm interference. These are reported results for this particular controlled lamp test—not independent measurements or a general benchmark. They make the example specific: the question is not merely whether the edited lamp looks plausible, but whether the relationships being protected still hold.

The workspace distinguishes learned hypotheses, deterministic facts, and source metadata. A component name carried in the source file is not, by itself, evidence that the system learned the object's meaning. Likewise, a proposed mechanism needs geometric support, and a successful edit needs checks tied to the constraints it was meant to preserve.

The lamp screenshots explicitly identify the current boundary: a controlled, generated lamp, with transfer to arbitrary imported assemblies still unproven. The earlier BRep workspace documents investigation on an Ergokiwi handle, including exact face queries and experimental operation-region hypotheses. Those hypotheses are not recovered CAD history, and the screenshot does not establish a completed edit.

The existing diagrams place the work in a larger CAD workflow: understand geometry, describe a change, preserve selected features, and validate the result. They illustrate the intended experience, not evidence that every illustrated capability is available.

The canonical roadmap records capability contracts, deterministic geometry support, human review, benchmark and adversarial testing, a real-BRep-part pilot, and Rhino and Grasshopper integration work. Its task-detail view makes the standard explicit: passing synthetic fixtures is not enough. The next research question is how reliably the same separation of interpretation, operation, and evidence holds across unfamiliar parts and assemblies.