An Iron Man–inspired holographic orb built with Next.js, Three.js, and MediaPipe hand tracking — control it with your bare hands through your webcam.
🔮 This repository contains the open-source interface for ULTRON, a real-time AI assistant and control system built and maintained by Ashish (@ashish7802).
🧑💻 GitHub: @ashish7802
demo.mp4
ULTRON Orb UI is a futuristic holographic interface inspired by sci-fi HUD design. It combines:
- a 3D orb environment built with Three.js
- gesture control via MediaPipe hand tracking
- browser-based voice interaction and local AI orchestration
- a cinematic, immersive dashboard experience
npm install
npm run devOpen http://localhost:3000.
Run npm run build:desktop to create installer/output/ULTRON-Setup.exe.
The installer creates Start Menu and desktop shortcuts and pins ULTRON to the
taskbar. The desktop app opens in its own window; it does not require a browser
window. It uses the laptop's front-facing webcam for hand gestures and requires
the Microsoft Edge WebView2 Runtime, which is included with current Windows
installations.
Building the Windows setup requires an unexpired, trusted Authenticode Code
Signing certificate with an accessible private key and SignTool from the Windows
in the CurrentUser\My certificate store. Set ULTRON_SIGN_CERT_SHA1 to its
40-character thumbprint, and optionally set ULTRON_SIGNTOOL to signtool.exe
and ULTRON_TIMESTAMP_URL to a trusted RFC 3161 timestamp service:
$env:ULTRON_SIGN_CERT_SHA1 = "<40-character-thumbprint>"
$env:ULTRON_SIGNTOOL = "C:\path\to\signtool.exe" # optional if already on PATH
npm run build:desktopThe build signs all unsigned executable and DLL payloads and the setup/uninstaller, timestamps and verifies each signature, and only replaces the installer after verification. Self-signed certificates are rejected and do not satisfy Smart App Control. Microsoft recommends Artifact Signing for Smart App Control compliance; a publicly trusted code-signing certificate is not a guarantee of Smart App Control's reputation decision. Cloud signing services must provide a SignTool-compatible certificate/key provider for this build flow; otherwise, provider-specific signing integration is required.
Voice chat uses the Windows default microphone, local Whisper transcription, and Ollama with dual-brain intelligent routing:
- Smart Brain (
qwen2.5:1.5b): Complex reasoning, multi-step problem solving, math logic, tool calling, and anti-hallucination. - Fast Router / Local Kernel: Instant (0ms) system actions (time/date, orb reset, diagnostics) and low-latency greetings (
qwen2.5:0.5b).
Pull both models before starting:
ollama pull qwen2.5:1.5b
ollama pull qwen2.5:0.5b| Input | Action |
|---|---|
| Drag | Spin the orb |
| Scroll / pinch | Zoom in & out |
Click GESTURES OFF (or press G) and allow camera access, then:
| Gesture | Action |
|---|---|
| Pinch (thumb + index) one hand and move it | Spin the orb |
| Pinch with both hands, spread apart / bring together | Zoom in / out |
| Key | Action |
|---|---|
G |
Toggle hand gestures |
V |
Toggle voice chat |
T |
Open the text prompt |
R |
Reset the view |
+ / − |
Zoom in / out |
Voice chat uses the browser's speech playback, the Windows default microphone, local Whisper transcription, and Ollama. Allow microphone access when prompted. If ULTRON does not hear you, select the intended input under Windows Settings → System → Sound → Input; a virtual audio device can be selected as the system default instead of the built-in microphone or headset. When voice mode is enabled, speak and watch the MIC meter; if it stays at 0%, check that input with Windows Settings → System → Sound → Input → Start test. Calling apps can override the Windows default, so select Microphone Array (Intel Smart Sound Technology) (or the mic you actually use) in each app's audio settings.
Install the local requirements once:
py -m pip install faster-whisper pywebview
ollama pull qwen2.5:1.5b
ollama pull qwen2.5:0.5bRun test and benchmark suites:
python benchmark_models.py
python tests/run_all_evals.pyULTRON.bat starts the local services and desktop window from the source
folder. When running the UI with npm run dev, start python stt_server.py
and make sure Ollama is running separately.
-
lib/intentRouter.ts— Dual-brain & hybrid router with zero-hallucination guardrails, routing instant commands, permanent memory ops, agent planning, and desktop tool calls. -
ultron_memory.py— Permanent SQLite memory engine with FTS5 virtual tables for BM25 semantic/keyword search, selective fact preservation, and memory lifecycle management. -
ultron_tools.py— Real computer control engine with system telemetry (psutil), app launch/close, file search/IO, developer commands, and confirmation token security gates. -
ultron_planner.py— Autonomous controlled agent loop: Understand$\rightarrow$ Plan$\rightarrow$ Execute$\rightarrow$ Verify$\rightarrow$ Report with step-by-step verifiable success criteria. -
launch_app.py— Production Windows supervisor with single-instance Named Mutex, auto-restart exponential backoff, structured JSONL logging, and clean process-tree termination. -
lib/orbScene.ts— Three.js holographic particle orb connected to real assistant states (idle,listening,thinking,executing,speaking,error,gesture). -
components/JarvisOrb.tsx— Interactive HUD interface with full-duplex barge-in voice interruption, sentence-by-sentence streaming TTS, and cancel/stop controls. -
stt_server.py— Local Whisper STT & unified Core API server for memory, tools, and planner execution.
Run the full evaluation runner to verify all 6 test suites:
python tests/run_all_evals.py| Suite | Status |
|---|---|
| Hybrid Intent Router | 100% PASSED |
| Permanent Memory (SQLite + FTS5) | 100% PASSED |
| Real Computer Control & Safety | 100% PASSED |
| Autonomous Agent Planner | 100% PASSED |
| Tool-Calling (Smart Brain 1.5B) | 100% PASSED |
| Anti-Hallucination & Factual Guardrails | 100% PASSED |
MIT
