NebiusCRACK THE CUBE← Back to the race
CRACK THE CUBE / SYSTEM ARCHITECTURE

THE RACE, UNDER THE HOOD.

Nebius Cloud coordinates the game. Token Factory hosts the models. Cloudflare connects the booth.

DEPLOYMENT SNAPSHOT23 OCT 2026
NEBIUS CLOUD · EU-NORTH1

From the booth screen to a verified cube

ONE CPU APP · THREE PARALLEL MODEL LANES
Crack the Cube deployed service architecture Visitors connect over HTTPS to the Cloudflare Worker at kube.curiousjaxx.dev, which proxies traffic to a managed Nebius HTTPS tunnel and the Serverless AI CPU app. The app makes three parallel Token Factory vision calls to Qwen, GLM and Nemotron. It runs a shared Soma solver, validates the results and streams race updates back to browsers. SecretStash injects credentials at startup. A persistent shared filesystem stores verified race records for replay and the daily leaderboard. NEBIUS CLOUD / jc-kubecon-demo / eu-north1 NEBIUS TOKEN FACTORY Managed vision inference Same image. Same voxel geometry. HTTPS+ SSE HTTP :8080 3 INDEPENDENT, PARALLEL CALLS Startup injection Read / write Booth browser3D cubes · race timersTHREE.JS / WEBGL PUBLIC ENTRY POINTCloudflareDNS · TLS · Worker proxykube.curiousjaxx.devsoma-kubecon-edge Managed HTTPS tunnelNebius endpoint ingress · encrypted public upstream SERVERLESS AI / CPU ENDPOINTThe race coordinatorNode.js · 2 vCPU / 8 GiB · serves the UI + API Signed sessions · parallel dispatch · shared Soma solverGeometry validation · event stream · daily leaderboard SecretStashProvider key · demo codeSession signing key Shared filesystem1 GiB · race records + replays/app/data/runs MODEL / 01Qwen3.8-27BVision input → solver tool call MODEL / 02GLM-5.3-FlashVision input → solver tool call MODEL / 03Nemotron 3.5 VLVision input → solver tool call
  1. 01 / THE BOOTH

    Visitor browser

    Three.js renders the cubes, timers and leaderboard. HTTPS requests and race events use the same public hostname.

  2. 02 / THE PUBLIC EDGE

    Cloudflare

    DNS and TLS for kube.curiousjaxx.dev. The soma-kubecon-edge Worker proxies requests and streams race events.

  3. 03 / NEBIUS CLOUD · EU-NORTH1

    Serverless AI CPU app

    A managed HTTPS tunnel reaches the Node.js app on port 8080. One 2 vCPU / 8 GiB container serves the UI, signs sessions, dispatches models, runs the shared solver, validates cubes and calculates daily bests.

    SecretStash → app: provider key, demo code and signing key, injected at startup.
    App ↔ shared filesystem: 1 GiB of persistent race JSON and replay data at /app/data/runs.
  4. 04 / NEBIUS TOKEN FACTORY

    Three models, in parallel

    Each gets the same image and known voxel geometry, then chooses a piece order for the shared solver. Results return to the app for checking and streaming to the browser.

    01 · Qwen3.8-27B02 · GLM-5.3-Flash03 · Nemotron 3.5 VL
Request / response + streamed eventsSecret injectionPersistent data
01 / INDEPENDENT CHOICES

Same input. Their own approach.

Each model chooses a search order for the same exact-cover solver. Matching solutions are allowed; an arrangement is never assigned to a lane.

02 / VERIFIED FINISHES

All 27 cells have to check out.

The backend validates piece shapes, rotations and overlap. Timers include model latency and solver execution; the visual replay happens afterward.

03 / PERSISTENT RESULTS

The best times stay on the board.

Verified race files survive app restarts. Daily model bests reset at Salt Lake City midnight. The provider key stays inside the backend.

THE DEPLOYMENT PATH / SEPARATE FROM LIVE RACE TRAFFIC

Private GitHub sourcegogococo/soma-kubecon
Manual Docker build + pushTested app container
Nebius Container RegistryPrivate image · daily-v1
Pull on endpoint startupsoma-kubecon-access-v2