Spatial Intelligence · Deeptech AI

Spatial intelligence for machines that reason in 3D.

A deeptech AI startup building AI that understands the structural and content of physical space — enabling machines to reason, build and operate in 3D. Unlocking new frontiers across gaming, architecture, and robotics.

Product

Spintel — 3D environments from text

Applied AI research

SCP — spatial context for agents

Fundamental AI research

CORD — object-centric world models

01 · Product — SPINTEL

Production-ready 3D

Spintel Scenes from natural language

A product suite for generating production-ready 3D environments from natural language descriptions.

Describe a building. Get a fully furnished, physics-valid scene — floor plan, furniture, wall fixtures, ceiling fixtures, small objects — ready for game engines, simulators, and renders.

Open Spintel

Loading scene…

Drag to orbit · Shift-drag to pan · Scroll to zoom

Fig. 1 · Scene Showcase Spintel · text → scene

A 13-room multi-functional building, built from scratch using a single text prompt. Floors, walls, fixtures, and furniture objects all generated by the Spintel text → scene pipeline.

02 · Applied AI Research — SPATIAL CONTEXT PROTOCOL

scp-agent · session 01

› How do I get from the entrance to the changing room?

→ From the entrance lobby, walk north and follow the corridor to the main junction. Turn west and take the third doorway from the south — between the orange noticeboard and the kitchen door. Total distance: 13.9 m.

› Turn off the lights in the lecture hall.

→ Resolved 6 lighting fixtures in lecture_hall_01. Issuing scp.actuate(zone="lecture_hall_01", target="lights", state="off") — done.

Live agent transcript. SCP exposes a building as a compact, LLM-interpretable spatial model — lightweight enough to run on modest hardware.

Open protocol · v0.1

SCP Spatial Context Protocol

An open protocol for AI agents to understand the physical world.

SCP enables LLMs to navigate buildings, identify locations, and control their environment in real time. Any space — generated, scanned, or hand-modelled — becomes a queryable, actuatable graph the agent can reason over.

Read the spec

03 · Fundamental AI Research — OBJECT REPRESENTATION LEARNING

Research programme

CORD Object-centric world modelling

Learning structured, holistic object representations from object detection.

World modelling requires object queries that capture what an object is, where it sits, what it canonically looks like, and how it relates to everything else in the scene — substantially richer than detection training alone.

Compositional object queries decomposes a detected object into four complementary channels — rich substrates for object-centric world modelling with true object and spatial reasoning.

— Get in touch

Building spatial intelligence for gaming, architecture, and robotics.

We're a small deeptech team. If you're building agents, simulators, or pipelines that need real spatial understanding — we'd like to hear from you.

[ LET'S CHAT ] →