Built in Rust · AGPLv3 Dual-Licensed

Koodesign

The AI-Native, Hardware-Accelerated Design Engine.

Bypasses the DOM entirely. Renders directly to a WebGPU surface — delivering consistent 120 FPS on canvases with 10,000+ nodes. Real-time CRDT multiplayer. Autonomous AI generation on the scene graph.

Read the Technical Blueprints
winitInput Events Tool FSMkd-client SceneCommandkd-core + Loro Tessellatelyon + DirtySet wgpuGPU Present kd-agent (AI) kd-server (WS sync) ≤ 2ms input → pixel
0 FPS
WebGPU Render Loop
≤0 ms
Input-to-Pixel Latency
0 MB
Static Binary (musl)
0%
Gross Margin
Scroll
Market Timing

Why Now?

Three technology convergences create the exact right window.

2024
WebGPU Ships Everywhere

Chrome, Firefox, Safari all ship WebGPU. The browser finally has a real GPU API — not a polyfill over WebGL.

2025
Loro CRDT Matures

Loro v1.0+ delivers production-grade collaborative editing with columnar encoding and sub-millisecond merges.

2025
LLMs Get Structured Output

LLM tool-use APIs enable reliable typed generation — AI can emit structured commands, not chat text.

Technical Architecture

Three Structural Moats

Not incremental improvements — fundamental architectural advantages that compound over time.

kd-render + kd-core

The Engine — WebGPU Renderer

Zero DOM. Zero Canvas 2D. The wgpu backend talks directly to Vulkan, Metal, DX12, or native WebGPU — delivering deterministic 120 FPS at 10,000 nodes with an 8.33ms frame budget.

Lyon CPU tessellation (zero T-junctions)
Bumpalo per-frame arena allocator
R-Tree spatial hit-testing (rstar)
SIMD affine transforms (glam)
// Render path: input to pixel in ≤2ms
winit::Event → ToolFSMSceneCommandDirtySet → tessellate → wgpu::present
kd-server

The Network — Lock-Free CRDTs

Loro CRDT — not OT, not polling. Conflict-free merge with deterministic convergence. Works offline, syncs on reconnect.

  • ≤100ms sync latency (localhost)
  • 50–500 byte deltas per operation
  • Columnar encoding (zero-copy)
  • Airplane-mode native
kd-agent

The Agentic Wedge — AI on the Scene Graph

Not a chatbot overlay. kd-agent connects to any LLM provider — cloud or local — via a unified streaming interface and injects typed SceneCommand[] directly into the live CRDT document — a first-class scene citizen.

≤8s
Wall-clock generation
≤2K
Input tokens/generation
Incremental
Streamed to canvas live
100% Rust

Memory-Safe by Construction

Zero buffer overflow CVEs. No GC pauses. No null pointer exceptions. The entire stack — GPU shaders to WebSocket parsing — is Rust.

Post-NSA advisory compliant — no managed runtime attack surface
Competitive Landscape

Koodesign vs. Incumbents

Architectural differences, not feature checklists.

Koodesign Figma Penpot
RendererWebGPU (wgpu)WebGL2 + CanvasSVG DOM
LanguageRust (native+WASM)C++→WASM + TSClojure + JS
Sync EngineCRDT (Loro)OT (proprietary)Server-auth
OfflineFull (CRDT native)LimitedNo
AI IntegrationScene graph nativePlugin overlayNone
Self-Host18MB binaryNoDocker (multi-GB)
LicenseAGPLv3 + CommercialProprietaryMPL-2.0
10K Nodes120 FPS / 8.33ms~30 FPSUnusable

Unified Command Architecture

Human input, AI generation, and network sync all flow through one type: SceneCommand

// Every mutation — human, AI, or sync — is a SceneCommand
enum SceneCommand {
    AddNode   { parent: NodeId, kind: NodeKind, frame: Rect },
    MoveNode  { id: NodeId, delta: Vec2 },
    SetFill   { id: NodeId, fill: Fill },
    DeleteNode{ id: NodeId },
    // ... 12 variants total
}

// Applied atomically to Loro CRDT → broadcast → retessellate
fn apply(doc: &mut LoroDoc, cmd: SceneCommand) → DirtySet<NodeId>
Enterprise Deployment

The Business Case

One binary. Zero dependencies. Air-gapped corporate networks. 95.6% gross margin at scale.

Air-Gapped On-Premise

A single 18 MB statically-linked musl binary serves the full application: HTTP UI, WebSocket sync hub, and AI agent stub. No Docker. No node_modules. No runtime dependencies. No outbound connections.

$ cargo build --release --target x86_64-unknown-linux-musl \
  --features "enterprise,agent"

# Output: single self-contained binary
-rwxr-xr-x 18M kd-server
Offline Ed25519 JWT License

Public key compiled into binary. No phone-home. No NTP. On expiry: graceful degradation to Free tier — never hard shutdown. Renewal: new JWT → env var → restart.

Unit Economics @ 1,000 Pro Seats

Monthly Revenue$20,000
Hetzner AX102 (bare metal)-$180
LLM inference (1K × $0.60)-$600
Storage + CDN + Infra-$105
Total COGS-$885
Gross Margin95.6%

No Kubernetes. No Terraform. Single Hetzner box: koodesign-server + Postgres 16 + Redis 7 + Caddy. Enterprise on-prem: COGS ≈ $0 → 98% margin.

AGPLv3 Dual-License Strategy

Full source under AGPLv3 — network use = distribution. SaaS competitors must open-source modifications or buy a Commercial License. The Elastic/MongoDB playbook: copyleft replaces the moat that proprietary code once provided.

CLARequired
Free (AGPLv3)
$0
3 collaborators · 50 AI tokens/day
Popular
Pro
$20/seat/mo
25 collaborators · 200K tokens/mo
Enterprise
$99/seat/mo
On-prem · SSO/SAML · Unlimited AI
Execution Plan

Roadmap

19-week engineering sprint across three phases.

P1
CURRENT — 8 weeks
Engine Chassis
Scene graph + SceneCommand
Lyon tessellation + wgpu init
Loro CRDT integration
120 FPS render loop
WASM browser build
P2
NEXT — 6 weeks
Real-Time Collab
WebSocket sync hub (axum)
≤100ms localhost latency
Conflict resolution rules
Snapshot persistence
500 concurrent connections
P3
PLANNED — 5 weeks
AI Agent Loop
Multi-provider streaming
Viewport context culling
Incremental scene generation
≤8s wall-clock generation
BYO model support

Built on proven foundations

wgpu lyon loro glam axum winit taffy tokio rstar slotmap multi-llm

Ready to discuss?

Whether you're evaluating for a grant committee, investment thesis, or enterprise deployment — let's talk architecture.