Skip to content
spatial brain · early accessworkbench / q-agent · build 4a7c

AI design-review agents for CAD assemblies

Quatrion turns complex CAD assemblies into a spatial reasoning layer that engineering teams can query for tolerance, manufacturability, simulation triage, ergonomics, safety, and design-review risks — with evidence tied back to geometry.

Built for CAD-heavy engineering teams where design reviews, simulation handoffs, and manufacturability checks still take days.

front · side · top → editable 3D solidmanifold · replayable
/ The problem

Engineering review is still trapped in tool handoffs

CAD, FEA, tolerance analysis, manufacturability checks, and compliance reviews often live in separate workflows. Every new question can mean another export, another mesh, another manual review, and another delay.

Before

Hours to days of preprocessing, handoffs, screenshots, and manual engineering review.

After

Encode the assembly once. Run multiple design-review agents against the same spatial representation.

·Initial wedge: first-pass design-review triage for CAD-heavy automotive, aerospace, industrial machinery, and robotics teams.
/ 01 · Workbench

Four surfaces. One workbench.

The shell every agent renders into — and the deployment model that lets your team run it inside your network boundary, against your own files.

/ 01 · 3d
early access

3D

Interactive CAD viewer for assemblies and parts — selection, sectioning, and spatial context shared with chat and runs.

Visuals
/ 02 · visuals
early access

Visuals

Charts, dashboards, and overlays that render directly on top of model and analysis data — no export round-trip.

/ 03 · vault
early access

Vault

Searchable, workspace-native repository of assets, templates, prompts, agents, recipes, and tools.

Runs
/ 04 · runs
early access

Runs

Reproducible analysis runs — capture inputs, parameters, outputs, and provenance, then replay or share them.

/ 01.1 · Architecture

Cloud-agnostic by design. No managed-runtime lock-in.

A polyrepo of independently deployable services behind a single workbench. Storage, secrets, identity, and job state are abstracted; nothing on the critical path requires a specific cloud provider.

┌─ q-agent deploy --targetsaws · gcp · azure · on-prem · local
· tenet.01

Polyrepo, not a monolith.

Independently deployable services across the workbench shell, agent runtime, CAD kernel, encoding, simulation, conversion, and platform layers — each on the compute profile it needs. CPU services bootstrap independently of GPU services.

· tenet.02

Cloud-agnostic adapters.

Object storage, secrets, identity, and job state sit behind small interfaces. The same Helm charts deploy against AWS, GCP, Azure, or on-prem Kubernetes — and swapping a provider is a config change, not a rewrite.

· tenet.03

Open agent runtime.

The reasoning layer is open and swappable, exposed through an open NDJSON endpoint. LLM provider, model, and tool surface are swappable via configuration — your reasoning layer is not bound to a managed cloud agent service.

· tenet.04

Local-first option.

Raw CAD can stay inside a workstation, on-prem, or private-cloud boundary. The workbench points at configured runtime endpoints; only configured queries, metadata, or approved artifacts cross that boundary. Air-gapped deployment can be discussed on request.

Supported deployment targets
·AWS EKS·GCP GKE·Azure AKS·On-prem Kubernetes·Local Rancher
/ 01.2 · Fit

Between your CAD stack and the decisions that come out of it.

Quatrion does not replace CATIA, NX, SolidWorks, Ansys, Abaqus, PLM systems, or engineering judgment. It adds a geometry-aware reasoning layer that helps teams triage issues earlier, preserve evidence, and accelerate review workflows.

/ Inputs

Your CAD & simulation stack

  • CATIA · NX · Creo · SolidWorks · Onshape
  • Ansys · Abaqus · Hyperworks · Star-CCM+
  • JT · STEP · GLB · IGES · STL
  • PLM systems of record
Remain the systems of authoring & release.
/ Quatrion

Spatial-reasoning substrate & agent platform

  • Spatial Brain & CAD kernel
  • Reference agents + agents your team authors
  • Workbench shell · provenance · replay
  • Cloud-agnostic, local-first option
Becomes the system of reasoning & evidence.
/ Outputs

Decisions, evidence, downstream handoff

  • Design-review findings with provenance
  • FEA prioritization & cycle triage
  • Manufacturing & DFM signals
  • Standards-referenced analysis runs
Remain the systems of decision & release readiness.
/ 01.3 · Evidence

Evidence, not opinions. Tied back to geometry.

Every finding the workbench returns carries its provenance — so a result can be reviewed, replayed, and trusted, not just read.

· 01

Replayable reasoning

Every run is captured so it can be replayed, shared, or compared to a prior run.

· 02

Geometry-linked evidence

Findings point back at the specific geometry references behind them.

· 03

Engineering audit trail

Inputs, tool chain, and outputs are recorded as structural provenance.

/ 02 · Agents

Design-review agents for real engineering workflows.

Grouped by the engineering decision they support — not by algorithm. Each group runs against the same encoded assembly, and your team can author more through the SDK without rebuilding the encoding, routing, or evidence layers.

/ Group 01

Design review

Catch geometric and assembly-level issues before they reach a release gate.

  • Tolerance verification
  • Surface continuity
  • Part consolidation
  • Clearance checks
/ Group 02

Simulation triage

Decide which parts and load cases deserve a full solver run first.

  • NVH screening
  • Buckling susceptibility
  • Stamping / forming risk
  • CoG migration
/ Group 03

Ergonomics & safety

Screen occupant and operator interactions against standard envelopes early.

  • Reach envelope
  • Vision obstruction
  • Head-impact zones
  • Human access constraints
/ 02.1 · Benchmark workbench

Reference agents demonstrated on the encoded benchmark assembly. They are the patterns your designers extend from via the SDK.

┌─ q-agent ls --reference10 ref · 1 sdk-slot
idagentquerymaps_tostatus
/01Dimensional tolerance verification"Which assembly joints exceed specified geometric tolerance?"DTS · GD&T · SAE J1100reference
/02Stamping strain risk"Where will this panel thin, wrinkle, or split?"Forming feasibility screenreference
/03NVH frequency screen"Does this panel land in idle, road-boom, or wind bands?"Modal density · 25–500 Hzreference
/04Panel buckling susceptibility"Which panels show low out-of-plane stiffness under expected loads?"Stiffness susceptibility scorereference
/05Class-A G2 continuity"Where do surfaces fail Class-A continuity?"G0 / G1 / G2 classificationreference
/06Head impact zone (GRRA)"Which interior points fall inside the 165 mm ball reach?"ECE R21 · FMVSS 201reference
/07Pillar vision obscuration"Where does the A-pillar block driver sightlines?"ECE R125reference
/08Manikin reach envelope"Can a 5th–95th percentile occupant reach the controls?"SAE J826 / J833reference
/09ICE → EV CoG migration"How far has the center of gravity moved on this platform?"Geometry-derived mass-envelope estimatereference
/10Part-consolidation candidates"Which parts could merge without changing form or function?"Constraint graph + BOMreference
/sdkYour next agent"define question · inputs · evidence"Domain SDK · tool registry · event bussdk
/ 03 · Benchmarkinternal · first-pass triage

One assembly, encoded once — queried by every agent.

In a 1,492-part automotive body assembly, Quatrion encodes the geometry once into a compact query cache, then runs every first-pass agent against the same spatial representation — no re-meshing per question.

┌─ q-encode baseline_silverado.glb1 cache · 10 agents
Source CAD
100–500 MB
catia · nx · step · 1,492 parts
encode · once
Spatial Brain™
core IP
< 5 MB
query cache · versioned · built 1×
query · repeatedly
10 reference agents · same cache
tolerancestamping strainNVH screenbucklingClass-A G2head-impactpillar visionreach envelopeCoG migrationpart consolidation+ your agent · authored via SDK
/ 01 · compression
~95%
typical compression of source CAD into the query cache
/ 02 · throughput
~100×
GPU batch compute vs CPU baseline
/ 03 · triage
~10×
faster first-pass design-review triage vs specialist tool handoffs
/ 04 · tolerance
≤1 mm
deterministic dimensional-tolerance measurement on encoded assemblies

Benchmark figures from a representative 1,492-part automotive body assembly. Intended for first-pass design-review triage — not a replacement for final engineering validation.

Read benchmark notes →
/ 04 · Method
Core IP

Encode once. Query repeatedly.

The Spatial Brain · three pieces of math, one substrate

Quatrion compiles a CAD assembly into a compact, queryable spatial layer before any solver touches it — every agent then inspects geometry, tolerances, surfaces, and clearances without starting from scratch each time. The math below is the public face; schemas, encoders, and benchmarks stay IP-protected.

/ Step 01 · the math

CAD becomes a compact geometry substrate.

Each part is encoded as integer-quantized position, orientation, and a shape descriptor. A 100–500 MB assembly collapses to a 1–5 MB cache that downstream agents and tools query directly — no re-meshing per question.

Compact spatial schema · quantized pose · learned shape codes over geometry features
/ Step 02 · the math

Shape behavior is captured as compact screening features.

For each surface, Quatrion computes geometry-derived screening features that help prioritize NVH, stiffness, and thermal-risk review before a full material-aware solver run. These features are intended for first-pass triage, not replacement of material-aware simulation.

Per-surface geometry-derived screening features · GPU batch compute · ~100× vs CPU baseline
/ Step 03 · the math

Fit and tolerance risk is prioritized for review.

Mates, contacts, alignments, and clearance targets are organized into reviewable relationships so fit drift can be ranked by practical engineering impact. The goal is to surface systematic tolerance issues clearly, instead of asking an engineer to inspect thousands of part-pair distances.

Geometry-linked tolerance prioritization · systematic-vs-local issue separation · review-ready interface ranking
/ 04 · Build

And three pieces of authoring, that your designers own.

Reference agents on this page were authored against the same SDK your team will use. You define what the agent answers, what geometry it reads, and what evidence it returns — the workbench renders the rest.

┌─ q-agent author --new-agentsdk · early access
$ step.01

Declare the question.

A small declarative spec: what the agent answers, what inputs it accepts, what evidence it returns. The workbench renders it as a first-class command, with the same provenance and replay surface every reference agent uses — no UI work required.

Agent manifest (YAML / JSON) · tool input/output schema · workbench command surface · permissions scope
$ step.02

Read directly from the encoded assembly.

Agents query the same encoded assembly that reference agents do — per-part position, orientation, shape fingerprint, and the constraint graph between them. No re-meshing, no second ingest of CAD, no bespoke geometry plumbing per agent.

Spatial Brain API · constraint graph · proximity graph · per-part feature vector · CAD kernel edit sessions
$ step.03

Return evidence, not opinions.

Every run captures inputs, the tool chain that produced the answer, and the geometry references behind each finding. Replay it, share it, or compare it to a prior run. Auditability is structural — your agents inherit it, they don't have to implement it.

NDJSON event stream · run provenance · viewer overlay primitives · replay and compare across runs
/ Connectors

Bring models straight from your source of truth.

Many engineering models never arrive as a local upload — they live inside cloud CAD, PDM vaults, and enterprise PLM. Connectors pull a selected revision or configuration directly into Quatrion as a derived model for viewing and spatial reasoning, while preserving the source item, revision, and BOM path behind every finding.

Provenance, not detached geometry

Every import carries a visible source summary — system, revision, configuration, and BOM path.

A derived model, clearly labeled

Imports open in the same viewer, runs, and evidence surfaces — labeled derived, not the authoritative file.

Credentials stay on the boundary

Connections reference credential profiles, never raw tokens, signed URLs, or source files.

Honest failure, not silent fallback

An unsupported source returns a typed, actionable blocker — never a silent mesh or fabricated claim.

Onshape · Cloud CAD · Part Studios & AssembliesAvailable
Autodesk · Vault + Platform ServicesPlanned
PTC · WindchillPlanned
Siemens · TeamcenterPlanned
SOLIDWORKS · PDM · 3DEXPERIENCERoadmap
Aras · InnovatorRoadmap

Connectors are read-only in v1. They import a selected revision as a derived model for review — they never check out, modify, or write back to your source system, and the source item/revision stays the authoritative CAD record.

Logos identify the source systems Quatrion connects to or plans to support, and reflect roadmap status, not availability. They do not imply affiliation, endorsement, or partnership — each mark is the property of its owner.

/ Team

Built by engineering and infrastructure operators

Quatrion is being built by operators with experience across cloud infrastructure, data platforms, AI systems, manufacturing workflows, and enterprise engineering environments.

The company is focused on one clear wedge: helping CAD-heavy engineering teams run first-pass design-review agents over complex assemblies, with evidence tied back to geometry.

$ q-agent request --access

Bring one assembly. Run one agent. See the evidence.

Start with a CAD assembly and one high-value review workflow. Quatrion will show how spatial agents can accelerate first-pass engineering triage while preserving evidence and provenance.

We review every request and follow up with qualified teams. The form helps us scope the right starting agent and assembly for your engagement.

We review every request and follow up with qualified teams.
Quatrion

A command-first desktop workbench for CAD-native engineering teams. Grounded answers from your own files, in your own environment.

Platform
AgentsWorkbenchConnectorsBuildDocs
Company
How we fitEarly accessContact
Resources
Docs
© 2026 QuatrionSpatial Brain™ is proprietary technology of Quatrion. All rights reserved.