Scan a room with your phone. Get a 3D Gaussian-splat world that Claude, ChatGPT, Gemini, and Copilot can see, measure, and reason about, all over MCP.
NVIDIA Inception member · AWS Activate · In production at rakuai.com
The problem
AI assistants are blind to physical space
Today's models read your code, your email, your files. They still can't see the room you're standing in.
0
rooms
have any 3D awareness in today's AI assistants by default.
3B+
phones
already pack the camera and the compute to capture one.
1
protocol · MCP
is the standard every modern assistant now speaks.
What RakuAI is
Phone → 3D world → AI sees it
A capture-first spatial AI platform built on Gaussian splats and MCP.
1
Capture
A guided phone walkthrough captures the space in 30 to 60 seconds. The PWA installs from any mobile browser and works offline.
2
Reconstruct
COLMAP → GLOMAP → Brush → splat-transform. A real Gaussian-splat .spz lands on CDN, with a metric scale solved from a credit-card reference.
3
Reason
The scene is exposed to LLMs over MCP. Claude, ChatGPT, and Gemini query geometry, run measurements, and run spatial reasoning prompts.
The stack
Three layers, all real today
In production at rakuai.com. 5 patents. 9,500+ API endpoints. A C++ engine, Python orchestration, and a phone PWA.
raku-runtime
C++ engine · MCP surface
A C++ spatial engine across 18 native DLLs. It exposes thousands of functions over MCP through 17 native tools, plus a smart-glasses and Android XR launchable web-app entry.
raku-api
Python backend · 9,500+ endpoints
FastAPI orchestration. The job pipeline, capture lifecycle, accounts, connect-flow, and GPU upload routing, with Redis-backed multi-instance state.
raku-capture PWA
Phone · offline shell · walk-the-room
An installable phone PWA. Guided capture, a live coverage heuristic, and a splat viewer on Spark and three.js. Hosted at rakuai.com.
The pipeline
GPU-accelerated reconstruction
The pipeline is GPU-bound the whole way, from camera-pose recovery through splat training.
COLMAP
CUDA feature extraction and matching
60–120s
→
GLOMAP / FastMap
Camera-pose recovery (CUDA)
30–90s
→
Brush
Gaussian-splat training (wgpu / Vulkan)
2–6 min
→
splat-transform
Compress to .spz (CPU)
5–15s
Target GPUs: L40S, A10, A100. Target time-to-splat is under 8 minutes per scan, at a cost under $0.50 per reconstruction.
Why now
Three waves landed in one window
Three independent shifts hit inside the same 18 months. None of them was true a couple of years ago.
MCP is here
Anthropic shipped the Model Context Protocol. Claude, ChatGPT, and Gemini all consume it. A standard way to hand an LLM spatial tools didn't exist before 2024.
Gaussian splats matured
Brush, Spark, three-gs. Real-time splat rendering on a phone is now a solved problem. The 3D representation just works.
Smart glasses launching
Meta Ray-Ban Display ships installable web apps. Android XR opens the same path. That closes the capture-and-view loop.
Reach
Every camera. Every assistant.
The glasses render, the phone or edge computes. One runtime, three device classes, every assistant.
Devices
Phone PWAToday
Meta Ray-Ban DisplayWeb-app path
Android XRSame web path
Assistants (via MCP)
ClaudeMCP native
ChatGPTMCP via connector
GeminiMCP via connector
Where it goes
Wedges into one platform
Once any room can be queried by an LLM, the use cases stack fast. The same runtime sits at the center of all of them.
Real estate & redesign
List a property in 5 minutes and let buyers walk it with an AI guide. Restage rooms with photoreal furniture through a Meshy.ai integration.
Insurance & inventory
A one-pass scan turns into itemized contents and measurements. For claims, moves, and downsizing, a 60-second capture replaces the clipboard.
Accessibility
Spatial assistance for low-vision and wheelchair users. "Is there a step at the front door? Will my chair clear the hallway?"
AR/VR creators
Splat assets for game engines, training environments, and OpenXR previews, from the same pipeline that powers consumer captures.
Traction
Already shipped
Production-grade code across three repos. Phase 1, Phase 2, and Waves 1 through 3 are merged to main.
RakuAI, LLC · Spatial AI for every assistant · Founded 2025 as BladeWireless (an MBA capstone at Saint Mary's College of California), renamed RakuAI in 2026. · kevin@rakuai.com · rakuai.com · github.com/RakuXR
