"A voice with hands."
ULTRON is my Iron Man-inspired AI: a holographic orb you can spin and zoom with bare-hand gestures through your webcam, a real-time voice agent you talk to like a person — and, unlike most voice assistants, it acts in the physical world. Ask it to unlock a phone, open YouTube, or run a search, and the screens in front of you light up and do it.
The demo reel blew up on release — 20,000+ likes within the first 5 hours of publishing.
The Idea
Voice assistants are everywhere, but they all live behind glass: they talk, they fetch information, and that's where it ends. I wanted the Tony Stark experience — an AI with a presence (a holographic orb that reacts to its own voice), a personality, and hands: the ability to reach out and operate real devices while you watch.
How It Works
ULTRON is three layers working together:
- 🔮 The Orb (open source): A holographic wireframe orb built with Three.js — layered shells, a glowing core, orbiting debris, drifting code fragments, and a cinematic bloom/chromatic-aberration pass. You can spin and zoom it by pinching mid-air, tracked through the webcam with MediaPipe hand landmarks.
- 🎙️ The Voice (proprietary): A real-time speech-to-speech agent with a low-latency WebRTC pipeline and a custom in-character audio treatment. The orb is wired to the live output audio, so it visibly listens, thinks, and "mouths" every word as it speaks.
- 🤖 The Hands (proprietary): A device-control layer that gives the agent command over a rack of Android devices. Fast scripted actions land within a conversational beat; anything else is handed to an autonomous agent that reads the screen, taps, types, and verifies its own work — step by step until the goal is done.
Open Source & What's Next
The orb UI — the full Three.js scene and the MediaPipe hand-gesture control — is open source:
- ultron-by-sagar-builds — run it locally with
npm install && npm run devand control the orb with your bare hands.
The voice agent and the device-control layer are proprietary for now — a public release is coming soon. Meanwhile, if you want a fully open-source voice agent you can run today, check out F.R.I.D.A.Y. — my Tony Stark-inspired (female-voiced) assistant with 1.6k+ stars on GitHub:
Tech Stack
- Interface: Next.js, Three.js (WebGL, post-processing), MediaPipe Tasks Vision (hand tracking)
- Voice: OpenAI Realtime over WebRTC, custom Web Audio processing (see friday for inspiration)
- Device control: MCP (Model Context Protocol), Python, and other tech (releasing soon - follow for updates)
