TileGraphAgent

The Problem

Industrial 3D is not just triangles.

Industrial models are heavy, messy, proprietary, and semantically rich. A pump is not just a mesh. A valve is not just geometry. A pipe segment belongs to a line, appears in a P&ID, connects to equipment, has isolation logic, and must map back to visual features in the viewer. LLM agents cannot reason safely from triangles alone.

They need deterministic access to:

stable object IDs engineering tags graph connectivity spatial proximity P&ID references datasheets work packages tile & feature mappings auditable tool results

Traditional 3D Viewer

renders geometry, selects meshes

shows raw properties only

disconnected from graph semantics

no deterministic agent interface

preserves industrial object identity

maps visual features to engineering objects

exports semantic graph relationships

provides MCP tools for agents

logs every action for auditability

Core Thesis

A shared identity layer across 3D, graph, and agents.

TileGraphAgent is built around one principle: every visual feature, graph node, spatial index record, and MCP result must resolve to the same stable engineering object identity. This is what allows an AI agent to answer engineering questions with evidence instead of hallucination.

identity_stackschema

object_idobj_pump_p_1001

tagP-1001

source_idsynth/area-a

revision_idrev_004

tile_idarea-a-equipment

feature_id1201

gltf_nodePUMP-P-1001

global_transformmat4[16]

aabb[min, max]

Engineering Object

glTF Nodegeometry

3D Tiles Featurefeature_id

Spatial Index Recordaabb

Neo4j Graph Nodecypher

MCP Resourceuri

Viewer Selectionhighlight

Platform Architecture

End-to-end industrial 3D × AI pipeline.

Nine deterministic stages take a raw plant model from ingestion to agent-controlled viewer actions — each stage produces inspectable, validated artifacts.

Demo Scenario

Ask engineering questions. Get visual, graph-backed answers.

Find all pumps connected to LINE-1001, show their isolation valves, isolate the affected system in the viewer, and explain the maintenance impact.

agent_response · evidence-backed

LINE-1001 is connected to pumps P-1001 and P-1002.

Isolation valves

XV-1001A upstream of P-1001
XV-1001B downstream of P-1001
XV-1002A upstream of P-1002

Viewer action

Isolated SYS-COOLING
Highlighted 2 pumps, 4 valves, 11 pipe segments

Maintenance impact

Shutting down LINE-1001 affects the cooling water loop in Area A.

Confidence: High for synthetic graph data. Real plant validation required before operational use.

audit_log.jsonjson

{
  "session_id": "agent_session_001",
  "tool": "query_connected_components",
  "input": { "object_id": "obj_line_1001" },
  "output_count": 17,
  "status": "success",
  "timestamp": "2026-05-29T10:32:18Z"
}

MCP Agent Bridge

Deterministic tools for engineering agents.

TileGraphAgent does not allow the LLM to guess engineering facts or directly manipulate the model. The agent must use schema-bound MCP tools and return evidence from structured results.

Agent rules

Never infer facts without tool evidence.
Always resolve tags to object IDs.
Always distinguish connectivity from proximity.
Never trigger viewer actions without object mappings.
Always include uncertainty.
Always write audit-friendly answers.

Knowledge Graph

Engineering semantics beyond the mesh.

The graph makes engineering relationships explicit, queryable, and inspectable by both humans and agents.

Relationship examples

PumpPART_OFSystem

ValveCONNECTED_TOPipeSegment

LineAPPEARS_IN_PIDPID

PumpHAS_DATASHEETDatasheet

FeatureMAPS_TOObject

ObjectHAS_BOUNDING_VOLUMEAABB

ValveISOLATED_BYLine

IssueAFFECTSEquipment

isolation_query.cyphercypher

MATCH (line:Line {tag: "LINE-1001"})-[:CONNECTED_TO*1..4]-(pump:Pump)
OPTIONAL MATCH (pump)-[:ISOLATED_BY]-(valve:Valve)
RETURN line.tag, pump.tag,
       collect(valve.tag) AS isolation_valves;

3D Tiles + glTF Identity Mapping

Every visible feature maps back to an engineering object.

TileGraphAgent uses 3D Tiles for scalable streaming, glTF/GLB for optimized geometry, and metadata mapping files to preserve the link between visual content and engineering semantics.

gltf_node.extrasjson

{
  "name": "PUMP-P-1001",
  "extras": {
    "object_id": "obj_pump_p_1001",
    "tag": "P-1001",
    "class": "Pump",
    "system": "SYS-COOLING",
    "line_refs": ["LINE-1001"],
    "feature_id": 1201
  }
}

output/tiles/tree

tileset.json

content/

area-a-root.glb

area-a-piping.glb

area-a-equipment.glb

metadata/

feature_table.json

object_properties.json

tile_feature_map.json

index/

spatial_index.json

reports/

validation_report.json

CesiumJS Viewer

Agent-controlled 3D viewer actions.

The viewer is not just a display surface. It becomes the visual execution layer for deterministic agent actions.

ws://viewer · commandjson

{
  "type": "highlight_objects",
  "session_id": "agent_session_001",
  "objects": [
    {
      "object_id": "obj_pump_p_1001",
      "feature_id": 1201,
      "color": "#38BDF8"
    },
    {
      "object_id": "obj_valve_xv_1001a",
      "feature_id": 1407,
      "color": "#F59E0B"
    }
  ]
}

Validation & Correctness

Built for correctness, not demo magic.

Industrial agent systems must be verifiable. TileGraphAgent includes validation across geometry, graph, tiles, MCP schemas, and viewer mappings.

validation_report.jsonjson

{
  "objects_total": 168,
  "tags_unique": true,
  "missing_geometry": 0,
  "missing_graph_nodes": 0,
  "missing_tile_mappings": 0,
  "invalid_bounding_volumes": 0,
  "status": "pass"
}

All checks passed · 168 objects

Get Started

Explore the architecture behind agentic industrial 3D.

TileGraphAgent bridges geometry, graph semantics, and AI agents — turning industrial 3D models into queryable, auditable engineering systems.

Read the Technical Plan View Demo Scenario Open GitHub Repository

Industrial 3D,
made agent-readable.